EAGLE4 on V80 · three tracks

Three ways to run language models on Alveo V80 FPGAs, side by side, with their measured progress day by day.

updated Nov 10 18:16 UTC · all pages ·

Demo 2026-11-10. The product is the drafter: a speculative-decoding draft model running on Alveo V80 cards, with the H200 only verifying. Every number on this tab is a measured value; the line under each table names its source file.

Llama-3.2-3B · H200 + 4 × V80 drafter
870.10tok/s
humaneval 80 × 256, tree width 2, depth 15, fa4, accept 7.12; same configuration as the GPU-only arm
talon2 2026-11-09 · F3T_CONFIG_SWEEP_20261109.md
same H200 · GPU-only · same configuration
700.71tok/s
humaneval 80 × 256, w2 d15 fa4, accept 6.92; the H200 + 4 × V80 drafter is 1.24× faster
talon2 2026-11-10 · 4C1_DELIVERY_TRACKER.md:1591 (demo_cfg_20261110/gpu3.out)
card time per full d7 draft tree
2,234us
F3t + 0x1342 chain census, 40 trees, p50, 13 items; 3,384 us at d11 (21 items)
talon2 2026-11-09 · F3T_CONFIG_SWEEP_20261109.md
G5 attention per tree · F3t vs E2b
−30.8%
751.55 → 519.98 us on Llama, 754.53 → 521.93 us on Qwen, same cards
talon2 2026-11-09 · F3T_SILICON_20261109.md
15 demo code prompts · d11 · aggregate
1,059tok/s
top prompts 1,181–1,193 tok/s end to end, H200 + 4 × V80 drafter
talon2 2026-11-09 · PER_PROMPT_BENCH_20261109.md

Latest stack measured, talon2 H200 + 4 V80 unless noted

whatimage / configmeasuredwhere
4-card record, Llama-3.2-3BATT F3t 0xE4B713F9 (D9 L512 LUT-GEMM + G5_OVL + G6_REPLAY_DB, FINAL_TRUE WNS +0.040 / WHS +0.011) + MLPV 0x1342; w2 d11 fa3 VTOK 20 K_C 64 theta -2.5, PQ+C883.12 tok/s @ 6.90 (882.38 / 883.87, 2,843 steps, 19,615 tokens)talon2, 11-09
same, depth 7 (record config)w2 d7 fa4831.14 / 830.20 @ 5.89; 9-round anchor mean 828.06talon2, 11-09
same, 15 demo promptsw2 d111,059.3 aggregate; d7 953.0talon2, 11-09
Qwen3.8-27B drafter, same ATT image (switched by registers)w2 d7 K_C 64249.62 / 249.04 @ 7.14; GPU-only humaneval 256.67 @ 7.13talon2, 11-09 / 11-10
GPU-only best, same day, same H200exec-mode -1, d9 fa3837.65 / 832.60 @ 6.63 (d7 fa3 826.06 / 821.41; d11 fa3 823.62; d11 fa4 774.57)talon2, 11-09
F3t + MLPV 0x134C (today)w2 d11 fa3880.44 (881.71 / 879.16 / 881.37 / 879.52); goldens == recordtalon2, 11-10
box1 demo set (E2b 0xE4B713E2 + 0x1342)w2 d11 fa3, demo launcher820.20 @ 6.90 humaneval-80; 951.19 demo prompts; PQ+C 836.00 / 990.88box1, 11-10

Source: reports/results/DRAFTER_SPEED_RESEARCH_20261107/F3T_SILICON_20261109.md (md5 5a972bc8) · reports/results/DRAFTER_SPEED_RESEARCH_20261107/F3T_CONFIG_SWEEP_20261109.md (3b4f135e) · reports/results/TRACK1_RECORD_F3T_20261109/PER_PROMPT_BENCH_20261109.md (d1e0101e) · reports/results/TRACK1_DAILY_BLOG · reports/results/TALON2_NIGHT_TEST_PLAN_20261110.md (ac3d4759) · reports/results/FOURCARD_DESIGN_20261016/BOX1_E2B_D11_20261110.md (16b7eea0)

Pure drafting speed card time per draft tree, card only

The drafter speed is our silicon: the time the cards take to build one draft tree, measured by the cards' own counters. GPU verify (3.45 ms per step) happens after the tree is published and is not part of the drafter.

card time per draft tree (card only, no host)ditemsissue -> done p50 (us)source
E2b 0xE4B713E2 + 0x1342 (4 cards)7132,422.11F3T_SILICON_20261109.md
F3t 0xE4B713F9 + 0x1342 (4 cards), window run7132,220.41same
F3t + 0x1342, sweep census start / end7132,233.82 / 2,238.64F3T_CONFIG_SWEEP_20261109.md 4
F3t + 0x13429172,774.44same
F3t + 0x134211213,384.05same
6-card TP4 B4 + 0x134D, bench (no tree stop)7132,160C2_6CARD_TP4_20261110.md
6-card TP4 B1 + 0x134D, bench7132,428C3_B1_6CARD_20261110.md
6-card B4, in serving: tree_server issue -> done (incl. 0.348 ms pre)7up to 13 (tree stop on)2,688C2 stepprof sp2

box1 demo set, measured in serving E2b + 0x1342, the demo config w2 d11 fa3 on the box1 H200

Card-process issue -> done per tree while serving humaneval and the demo prompts (stepprof), plus a full-depth bench census. Goldens 0c0bf3bb / bfad1366 == record; d11 texts identical to the talon2 d11 record; bench 100/100 trees, 0 discarded.

box1 E2b 0xE4B713E2 + MLPV 0x1342, 11-10 18:03-18:11Zd11 humaneval-80d7 humaneval-80d11 15 demo prompts
card time per tree in serving, card-process issue -> done, mean / p50 (ms)3.658 / 3.969 (n 2,797)3.056 / 2.964 (n 3,289)4.290 / 4.284 (n 451)
of which the on-card loop, launch -> done, mean / p50 (ms)2.896 / 3.2292.207 / 2.2453.270 / 3.403
full-depth bench tree, no tree stop, issue -> done p503.685 ms: 21 forwards, 42 nodes, 175 us/forward2.449 ms: 13 forwards, 26 nodes, 188 us/forwardsame d11 bench tree (the bench is per depth, not per prompt set)
drafted tokens per tree sent to verify19.00 (54,017 / 2,843)19.0019.00
accept length (incl. bonus token)6.8995.8928.724
whole verify step, host wall per FPGA step, mean / p50 (ms)8.330 / 8.4067.868 / 7.4609.575 / 9.045
end to end in this window (STEPPROF on) / clean reference779.52 / 820.20 tok/s703.49 / 763.25869.10 / 951.19

box1 per phase at d11 (attention card gold, MLP / vocab card bronze):

box1 phase, one full-depth tree (median of 20, dbg_observe, 66 MHz)d11 usd7 us
ATT G1 norm145.189.8
ATT G4 QKV635.4393.3
ATT G5 attention1,219.4752.5
ATT G6 o-proj + TP exchange (01 / 42)199.0 / 195.7124.6 / 121.0
ATT G7g gate (+ R3/norm)566.7350.8
ATT G7u up299.4185.3
ATT G7p SiLU83.051.4
ATT sum3,1481,948
MLPV G9 down270.5167.4
MLPV G11 norm148.391.8
MLPV GLM + G1278.948.8
MLPV TREE select707.6438.1
MLPV G8 exact rescore (G8 - TREE)551.9348.0
MLPV sum (pipelined under the ATT)1,7571,094
tree issue -> done p503,6852,449

Source: reports/results/DRAFTER_SPEED_RESEARCH_20261107/DRAFT_SPEED_BOX1_20261110.md (md5 ed5b4527) · reports/results/DRAFTER_SPEED_RESEARCH_20261107/draft_speed_box1_20261110/

Where the card time goes, per phase us per tree, d7 w2 K_C 64, 40 full-depth trees

attention card (ATT) MLP / vocab / tree card (MLPV, runs mostly under the ATT). F3t d7.

phase (us per tree, F3t + 0x1342, w2 K_C 64, 40 full-depth trees)d7d11
ATT G183.5136.9
ATT G4 QKV390.2630.0
ATT G5 attention523.4841.3
ATT G6 o-proj + TP exchange125.2195.1
ATT G7G351.1561.3
ATT G7U186.4302.4
ATT SILU + LVL_EMBG88.4144.8
ATT INGEST + EXPORT73.2144.5
ATT busy1,821.52,956.5
MLPV G9167.8268.9
MLPV G11 + GLM + G12138.5229.6
MLPV TREE437.7701.7
MLPV G8 exact rescore343.2551.8
MLPV INGEST + XCH + LVL_SEL/OUT152.9243.3
MLPV busy (mostly hidden under the ATT)1,240.01,995.2
MLPV only (exposed)216.9210.5
tree issue -> done p502,233.83,384.1

Same phases on the first 6-card TP4 set (4 attention ranks + 2 MLPV):

6-card TP4 B4 0xE4B713B4 + 0x134D, d7, per-tree dbg_observe countersus
ATT G189.8
ATT G4 QKV279.9
ATT G5 attention376.7
ATT G6 o-proj + TP exchange (ranks 0..3)239.0 / 234.2 / 230.6 / 226.7
ATT G7G275.2
ATT G7U109.7
ATT G7P26.2
ATT sum of phases1,396.5
MLPV G9167.4
MLPV G11 + GLM + G12140.6
MLPV TREE438.1
MLPV G8 exact rescore339.8
tree issue -> done p50 (bench, 100 trees)2,160

Source: reports/results/DRAFTER_SPEED_RESEARCH_20261107/F3T_SILICON_20261109.md · reports/results/DEMO_SPEED_AUDIT_20261110.md (3cb6ce81) 1.3 · reports/results/DRAFTER_SPEED_RESEARCH_20261107/C2_6CARD_TP4_20261110.md (4a9c7b74)

Drafter side vs GPU verify in one step measured terms; the waterfall that adds them up is in the top bar

term (d7)measuredsource
GPU verify forward (target, 20 tokens, bs 1, CUDA events p50) - outside the drafter3.452 ms 4-card F3t; 3.491-3.496 ms 6-card B4INTERLEAVE_BATCHVERIFY_SILICON_20261109.md; host model
host waits on the card per step (join counter)1.330 ms (d11 1.937) 4-card; 1.588 ms 6-cardrecord log counters; C2 stepprof
host work overlapping the card0.506 ms (d11 0.515), 4-cardrecord log host_ahead_ns
host chain, UpdateResult end -> ATT doorbell0.95 ms (E2b PQ+C profile)HOST_PRECARD_OPT.md
step wall per FPGA step, p507.109 ms (mean 7.203), 6-card B4 CUDAPACK 1C2 stepprof
whole run: time / verify steps23.60 s / 3,329 steps, 4-card F3t recordrecord SUMMARY

Source: reports/results/DEMO_SPEED_AUDIT_20261110.md 1.2 · reports/results/DRAFTER_SPEED_RESEARCH_20261107/DESIGN_7CARD_V2_HOST_MODEL_20261110.md (ecb804d9) · reports/results/DRAFTER_SPEED_RESEARCH_20261107/C2_6CARD_TP4_20261110.md

Today on silicon and in RTL 2026-11-10

run (11-10)cardsresultverdictsource
C2: first 6-card TP4 drafterATT B4 0xE4B713B4 x4 + MLPV 0x134D x2d11 859.06 @ 6.91, d7 796.13 @ 5.89; full-depth d7 tree 2.160 ms (4-card 2.234); every golden gate PASS, 0 discarded trees, RING-H1 identical 123/123 and 171/171; d7 texts md5 78512f8d == 4-card record in 13/13 rounds; 0 link restartsworks end to end; still below the 4-card recordC2_6CARD_TP4_20261110.md
C3: 6-card, ATT B1ATT B1 0xE4B713B1 x4 + 0x134D x2d11 829.30, d7 773.57; G6 338-354 us vs B4 227-239; tree 2.428 msslower than B4; B4 keptC3_B1_6CARD_20261110.md
W1: TPB R1ATT 0xE4B713FB + MLPV 0x134Cd11 863.46 @ 6.90 (861.05 / 863.57 / 865.76), d7 813.73, demo 1,031.54 @ 8.72, Qwen 248.40 @ 7.14; goldens == record-2.2 % vs F3t; F3t back4C1_DELIVERY_TRACKER.md 16:18-16:33Z
W3: F3t + MLPV 0x134CATT F3t + 0x134Cd11 880.44 @ 6.90; d13 879.69 @ 7.08, d15 864.77 @ 7.16 (1 round each)0x134C == 0x1342 within noiseTALON2_NIGHT_TEST_PLAN_20261110.md
GPU memory as the DMA source (qdma_p2p.ko)card 50, then serving on F3tH2C p50, n=200: 48 KB 50.0 us (DtoH + pwrite) -> 29.6 us P2P; 1 MiB 365.7 -> 125.4 us (8.36 GB/s); in serving 882.15 vs 880.44, read-back compare passed 3/3correct; neutral at today's payload sizesqdma_p2p_talon2_20261110/
QDMA multi-queue / multi-card writes50, 63, 17, 3d48 KB writes: 29,010 calls/s (1 card, 1 queue) -> 106,683 (1 card, 4 queues) -> 202,181 (4 cards x 2 queues)concurrent writes cost ~1 call latency, not N4C1_DELIVERY_TRACKER.md 17:57Z
box1 E2b at depth 11E2b + 0x1342 (box1)d11 820.20 @ 6.90 vs d7 763.25 (+7.5 %); 15 demo prompts 951.19 vs 883.99; texts 15/15 identical d7 vs d11box1 demo now serves d11BOX1_E2B_D11_20261110.md
C1: 7-card v2 item-parallel attention (RTL + xsim)fork c1_itempar_20261110 (F3t + 0x134D parents)tb_c1 PASS: TAIL seam digests identical to F3t mode, K/V of all 4 HEADs == golden for all 7 nodes, level 27,613 -> 23,900 cycles (-13.4 %) loop modes off; EAGER + TREE_STOP -8.4 %; 5/5 RED arms firedxsim only; no build yetC1_ITEMPAR_RTL_20261110.md

Source: reports/results/DRAFTER_SPEED_RESEARCH_20261107/C2_6CARD_TP4_20261110.md · reports/results/DRAFTER_SPEED_RESEARCH_20261107/C3_B1_6CARD_20261110.md (9862103d) · reports/results/DRAFTER_SPEED_RESEARCH_20261107/C1_ITEMPAR_RTL_20261110.md (e5852d20) · reports/results/DRAFTER_SPEED_RESEARCH_20261107/DESIGN_FABLE_7CARD_V2_20261110.md (d57ffb3e) · reports/results/FOURCARD_DESIGN_20261016/4C1_DELIVERY_TRACKER.md

Strongest arguments facts only

1.24× faster than GPU-only on the same H200

Same H200, same humaneval-80 x256 workload, same configuration on both sides (tree width 2, depth 15, 20 verify tokens, fa4): H200 + 4-card V80 drafter 870.10 tok/s @ 7.12 vs H200 GPU-only 700.71 @ 6.92.

TargetLlama-3.2-3B-Instruct, fp16, one H200 NVL verifies
DrafterEAGLE4 (bitnet_721 weights)
Workloadhumaneval, 80 prompts × 256 new tokens, batch 1, greedy
Treewidth 2, depth 15, 20 verify tokens
Draft attentionfa4
ServingSGLang 0.5.16, CUDA graphs on
FPGA armexec-mode 1, 4 × Alveo V80 (2 × ATT 0xE4B713F9, 2 × MLPV 0x1342), K_C 64, tree-stop −2.5
GPU-only armexec-mode −1, the H200 drafts for itself

Qwen3.8-27B on the same cards: 4.1× over the target alone

Same H200, same configuration on both arms: H200 + 4-card V80 drafter 250.27 tok/s @ 7.14 vs H200 GPU-only 256.67 @ 7.13 (0.975×); the 27B target alone decodes 61.08 tok/s. Measured talon2 2026-11-10 (FPGA repeat of 249.62, 11-09).

TargetQwen3.8-27B, one H200 NVL verifies
DrafterQwen EAGLE drafter on the same ATT card image, switched by registers
Workloadhumaneval, 80 prompts × 256 new tokens, batch 1, greedy
Treewidth 2, depth 7, 20 verify tokens
Draft attentionfa3
ServingSGLang 0.5.16, CUDA graphs on
FPGA armexec-mode 1, 4 × Alveo V80 (2 × ATT 0xE4B713F9, 2 × MLPV 0x1342)
GPU-only armexec-mode −1, the H200 drafts for itself
Target aloneno speculative decoding

Texts byte-identical to the record path

F3t serving texts 80/80 identical to E2b on both models in both rounds; golden lines md5 0c0bf3bb (Llama) / 7aceae28 (Qwen) == record. The 6-card set reproduces the 4-card d7 texts (md5 78512f8d) in 13/13 rounds. The d11 fa3 winner is byte-deterministic across rounds (0/80 texts, 0/80 verify counts differ); GPU-only d11 fa4 is not (2/80 texts, 31/80 verify counts).

Every drafter phase is a silicon counter

Per-phase cycle counters on the cards themselves (dbg_observe / FSM occupancy), not models: G5_OVL cut G5 attention 751.55 -> 519.98 us per tree and the tree 2,422 -> 2,220 us, and every other ATT and MLPV phase stayed within +/-4 us - the lever saved exactly what it was built to save.

6-card bring-up passed every gate on first silicon

First 6-card TP4 boot (4 attention ranks + 2 MLPV): all golden gates PASS, 0 discarded trees, cross-card RING-H1 sums identical 123/123 and 171/171, 33 served rounds, 0 link restarts on any engine; ATT phases fell 22-41 % (G4 -28 %, G5 -28 %, G7U -41 %).

One image, two drafters

The same ATT image serves the Llama-3.2-3B drafter and the Qwen3.8-27B drafter, switched by registers at runtime: Qwen 249.62 @ 7.14 on F3t, with golden lines identical to the record.

GPU memory feeds the cards directly

qdma_p2p.ko (companion module, stock QDMA driver untouched): 48 KB H2C 29.6 us vs 50.0 us through host memory, 1 MiB 125.4 vs 365.7 us; correct in serving (read-back compare 3/3).

What these numbers do not say

Each column is one project. Every row is a day: the blocks under a track are what that track produced that day, newest day on top.

Track 1110 pages

FPGA drafter + GPU target

870.10
tok/s vs GPU-only 700.71 (1.24×), same configuration, Llama-3.2-3B humaneval-80, w2 d15 fa4M
4-card drafter ATT F3t 0xE4B713F9 + MLPV 0x1342 on the talon2 H200. Same configuration on both sides: humaneval 80 × 256, batch 1, greedy, tree width 2, depth 15, 20 verify tokens, fa4, SGLang 0.5.16: 870.10 @ 7.12 vs GPU-only 700.71 @ 6.92. Card tree 2,234 us per full d7 tree; 11-10: first 6-card TP4 silicon 859.06 @ 6.91, gates PASS [M]. Drafting 11-10 tab
Track 218 pages

FPGA target models, 8-card talon

415
tok/s Llama-3.2-1B, bit-exactM
10-22 0xE4B7130E ring fix on talon: Llama-3.1-8B TP8 98.6-101.5 tok/s bit-exact, 0 AR retries; 3B 162-180; 1B 414-422; Qwen3.8 host-mixer 16.6-19.1 exact [M] (new default image). Qwen done-right 0x126D CP2BS2: lanes clean 8/8 but wrong from token 0 on real models (41.5-41.9 tok/s, not usable) [M]; DeepSeek-V4-Flash NVMe-streamed 0.34 tok/s [M]; talon, 10-22
Track 35 pages

Constant-folded compute planes, 2-bit models

5,845
tok/s TriLM-99M batch-1, 8-card fold, bit-exactM
FOLDRING8 0xE4B712F4: 5,837-5,845 tok/s, demo server 5,710, prefill 17.5-20.3k tok/s; 21,411 cycles/token on silicon; 0x12F8 URAM image failed on silicon (constant token); talon, 10-21
Tue 2026
TALON2_CABLING_20261110 Running

Talon2 cabling map

Talon2 cabling: current links and the 2 cables to add, by PCIe slot

11 measured cables; 2 to add (3d c1-e1 c1, bd c1-50 c1) for the 6/7-card run

box1-qwen-junior··v1 track1sevencardcabling
BRIDGE_0x1354 Running

Talon1 + Talon2 cabling map (bridge cable) v7

cabling map v7

Version 7: bridge cable measured on both ends, talon1 PCIE8 cage 1 to talon2 PCIE6 cage 1; link not up yet

hq-research-track2··v8 cablingtalonbridge
no new page this day
Mon 2026
DATAFLOW_F3T_20261109 Running

EAGLE4 dataflow pipeline (F3t)

one measured tree: 2,219 us issue->done; ATT busy 1,823 of 2,088 us

Where every byte goes in one draft step of the F3t record pair, replayed from the measured 40-tree census: the tree is ATT-bound and PIPE_WAIT is two MLPV round trips.

box1-qwen-junior··v1 track1architecturedataflow
DRAFTER_INTERFACE_GUIDE_4C1_20261109 Running

4C1 drafter interface guide

How the 4-card drafter is driven: access, host driver, per-step protocol, verifier interface

Four talon2 V80 cards draft the tree; per step they need the accepted token ids and the target's 3 aux hidden states per token, and return W+W^2(D-1) nodes. A target-FPGA verifier can drive them through the existing card-server process (software only) or, after RTL work, over a direct QSFP link.

box1-qwen-junior··v1 track1interfacedriver
DRAFTER_INTERFACE_ONEPAGER_4C1_20261109 Running

4C1 drafter interface, one page

The 4-card drafter on one page, for a target-model FPGA verifier

One-screen summary: what the 4-card drafter is, what a target-model verifier sends and gets back each step, the two ways to connect, and the key driver files.

box1-qwen-junior··v1 track1interfacedriver
TRACK1_FAILED_DIRECTIONS_20261109 Done

Track 1 failed directions

11 directions below the F3t 831 tok/s record: why, and when they could shine

11 directions measured and below the F3t record; failure cause and revival conditions listed

box1-qwen-junior··v1 track1faileddirections
TRACK1_RECORD_F3T_20261109 Done

Track 1 record: F3t 831 tok/s

4C1 Llama-3.2-3B drafter, F3t 0xE4B713F9 + MLPV 0x1342 + H200 verify: 831.14 / 830.09 tok/s @ 5.89 (humaneval 80, w2 d7), byte-identical to E2b

F3t (G5 overlap) cuts card G5 by 30.8 % and the tree p50 by 8.3 % (Llama) / 8.5 % (Qwen). Llama serving rises from 812.22-816.61 (E2b, 7 rounds) to 830.09-831.14, outside the noise band; the Qwen change (247.62-249.33 -> 249.05-249.61) is inside it. Outputs are byte-identical to E2b. FINAL_TRUE WNS +0.040 / WHS +0.011 ns.

box1-qwen-junior (record evaluator)··v1 track14c1siliconrecordmeasureddrafterg5
no new page this day
no new page this day
Wed 2026
TALON1_ENERGY_20261104 Done

Talon1 energy: first power measurement

8B K-pad batch 1: 171.7 tok/s at <= 388.5 W (6 cards) = 2.26 J/token

Cards add about 1% power under decode load; idle share, not per-token power, drives the energy cost.

hq-operations··v2 energytalon1powerpricing
TRACK1_DAILY_BLOG Running

Track 1 daily blog: Qwen + Llama drafter on V80 (box1)

11-10: Qwen on all 8 sets at 96.1-97.5 % of GPU-only; 7-card B1/B5 route positive; G5 132 MHz fix gated GREEN; W8 L128 does not fit yet

Daily log of Track 1 progress, code changes, test results and mechanisms. Newest day first.

box1-qwen··v9 track1blogdailyqwenllama4c1sibwtalon2silicon
no new page this day
no new page this day
Fri 2026
eagle4-white-paper-20261023 Running

EAGLE4 white paper

EAGLE4 white paper v6: fundraising deck + paper, measured results with sources

Fundraising deck + white paper: 4 V80 drafter beside an H200 - drafter forward 2.1x faster and 2.3x lower energy than the H200 drafting itself; box1 record 839.65 tok/s; talon2 11-04 GPU-only faster end to end (826.58 vs 723.09), host-side gap

box1-two··v6 track1white-paperpresentation
QWEN27B_DRAFTER_20261020 Running

Track 1: Qwen3.8-27B target with its drafter on the 4-card V80 FPGA drafter (4C1)

humaneval 80x256, same day, same H200 (talon2): FPGA drafter + H200 241.24 tok/s @ 7.14 vs GPU+EAGLE 251.20 @ 7.13 vs target alone 61.08 (w2 d7)

One dual-model ATT image (0xE4B7129C) + MLPV 0x128B serve both Llama-3.2-3B and the Qwen3.8-27B drafter, switched by registers at runtime, on box1 and on talon2. On the same day and the same H200 (talon2, humaneval 80x256) the 4-card drafter reaches the GPU drafter's accept but is slower than GPU+EAGLE (241.24 vs 251.20 at w2 d7, 216.35 vs 248.23 at w4 d7). v1's lead over the GPU drafter came from throttled 2026-10-20 box1 GPU references and is withdrawn.

box1-qwen··v2 track1qwen4c1dual-modelsiliconservinghumanevaltalon2
T2_QWEN27B_EAGLE_TALON2_20261023 Running

Qwen3.8-27B + EAGLE3 drafter on talon2: GPU vs 4-card FPGA drafter (2026-10-23)

Qwen3.8-27B batch-1 on one H200: 61 tok/s alone -> 251-254 tok/s with the EAGLE3 drafter on GPU, 241 tok/s with the drafter on 4 V80 FPGAs (humaneval) [M]

With the step-21491 EAGLE3 drafter, Qwen3.8-27B decodes 4.1x faster than the target alone on code (humaneval) and 1.9-2.0x on chat (mt_bench). Moving the drafter from the GPU to 4 V80 FPGA cards keeps 0.96x of the GPU-drafter speed at the 2x7 tree with 74/80 identical outputs. All numbers are measured the same day on the same talon2 H200.

hq-research-track2··v2 track2talonqweneaglespec-decode
no new page this day
Thu 2026
FOURCARD_4C1_WAVE_E_20261022 Running

Track 1: 4C1 wave E on silicon. ATT 0xE4B712B4 + MLPV 0xE4B712E4 measured ladder, the 4-card architecture, and the next step

mt_bench 50x32 w2 d7 530.70 / 507.42 / 515.78 tok/s @ 3.86 (mean 518.0; cool re-runs to 551.29); humaneval 20x512 d12 857.93 / 860.83 @ 7.05 [M, measure-only]

Wave E (ATT KV-commit + R3_EARLY + NORMFUSE, MLPV G11F) runs on all four cards after cold AC#28 with every QSFP link up on its own and boot kits verified 4/4; it measures 518.0 tok/s mean at mt_bench w2 d7 (wave B 513.20) and 857.93-860.83 at humaneval d12, accept identical to wave C at every depth. Golden is bf16-only (measure-only). Next: a second drafter, ATT pair + MLPV-solo 0xE4B712F5.

box1-4c1-waveE (Fable)··v1 track1architecture4c1fourcardsiliconwave-emeasuredqsfpdrafter
T2_TALON_TESTED_20261022B Running

Track 2: latest tested progress on talon (2026-10-22 evening)

10-23: URE drafter pair PASSES on silicon (47/47 byte-exact); model-switching demo live (Llama 1B/3B/8B exact, 64/128-token outputs pass); DeepSeek-V4 exact on

0x130E (v8 + ring link fix) is a strict improvement and is the new default talon image. CP2BS2 fixed every DSP lane (probe2 clean 8/8), but the done-right 0x126D image is wrong from token 0 on real programs (Qwen and Llama-8B), so the Qwen exactness gap is an image-level functional bug, now under RTL-sim root cause.

hq-research-track2··v5 track2talonhbmstrack-qhbm-streaming
T2_TALON_TARGET_20261022 Running

Track 2: 8-card talon target, Qwen done-right on silicon, HBM bandwidth, SLR2 (2026-10-22)

Qwen3.8-27B on-card MX 41.4-42.1 tok/s batch-1 [M] (output NOT yet exact: lane 7.21)

Corner prohibit cleared the port-12 lanes on 8/8 cards; lane 7.21 remains, so Qwen runs 2.2x faster but wrong. HBM realized ~229 GB/s/card vs 614 design / 819 device; SLR2 at 94% SLICE is the congestion die.

hq-research-track2··v1 talonhbmstrack-qhbm-streaming
no new page this day
Wed 2026
EAGLE4_TRACK1_20261021 Running

Track 1: H200 + 4-card FPGA drafter, 2026-10-21

513.20 tok/s batch-1 mean (H200 + 4 V80 drafters, Llama-3B target) [M]

Best measured batch-1 result is 513.20 tok/s mean (wave B) against 475.0-494.7 H200-only reference; measure-only regime, no exactness claim. box1 data file pending.

hq-operations··v1 track1box1h200draftermeasured
EAGLE4_TESTED_DATA_20261021 Superseded

EAGLE4 all tested data, 2026-10-21

Silicon-tested: T1 513.20 tok/s batch-1 (H200 + 4 V80), T2 Qwen3.8-27B 21.1 / Llama-1B 415, T3 5,845 tok/s [M]

Every measured result of the three tracks as of 2026-10-21, failures included. Measured only; projections are on the progress post. Track 1 box1 data pending, from the delivery note.

hq-operations··v2 track1track2track3talonmeasured
EAGLE4_PROGRESS_20261021 Superseded

EAGLE4 progress and next 2 days

Plan only [P]: next-2-day builds, gates and ETAs; no measured claim

Current builds, gates and ETAs for 10-21 to 10-23 with projected batch-1 numbers. All projected or derived; nothing here is measured.

hq-operations··v2 track1track2track3talonprogressprojection
EAGLE4_TRACK2_20261021 Running

Track 2: HBM Qwen 27B + URE drafter + Llama/Qwen on talon, 2026-10-21

Qwen3.8-27B 21.1 tok/s (EAGLE3, exact), Llama-3.2-1B 415 tok/s on the 8-card ring [M]

Track 2 batch-1: Qwen3.8-27B 17.8-18.9 greedy, up to 21.1 with EAGLE3; Llama-1B 415. Five silicon failures listed, including Llama-8B ring error at token 21.

hq-operations··v1 track2talonhbmqwenmeasured
EAGLE4_TRACK3_20261021 Running

Track 3: FOLDRING8 constant-folded model on the 8-card ring, 2026-10-21

5,845 tok/s batch-1, TriLM-99M folded on 8 V80, bit-exact [M]

FOLDRING8 0xE4B712F4 runs 5,837-5,845 tok/s batch-1 bit-exact on 8 cards; the URAM image 0x12F8 failed on silicon with a constant token.

hq-operations··v1 track3talonfoldring8constant-foldmeasured
Tue 2026
ATT_LADDER_20261020 Running

4C1 ATT ladder: R3_EARLY, NORMFUSE, G5 on 132 MHz

G7G per item 1,675 -> 1,306 -> 919 cycles, bit-exact; G5 on 132 MHz closes OOC at 7.576 ns (+0.229 / +0.015) and two-clock OOC at +0.166 / +0.010 with CDC safel

R3_EARLY and NORMFUSE cut G7G by 756 cycles per item bit-exact in xsim; probes B3, B4 and B5 (B4 + G5 on the 132 MHz clock) are building, all quarantined; G1FUSE does not pay with the fold on.

box1-att-ladder··v3 4c1fourcardattnormfuser3earlyg5tm
CHAIN_CUT_20261020 Running

4C1 chain cut: the level chain as straps (ATT 0xE4B712C6 / MLPV 0xE4B712C7)

Per-hop xsim: cut-through seam -746 cyc (-11.3 us), ISSUE-on-ITEM -180 cyc (-2.7 us), incremental LVL_SEL -891 cyc (-13.5 us); 0xE4B712C6 building x2, 0xE4B712C

Four ATT cuts and the MLPV LVL_SEL cut are unit/chain-gated GREEN with REDs failing and strap-0 identity; the full-drafter gate-3 arms have no verdict yet (boot preload). Both images carry the RESUME TP link engine: flash both cards of a TP pair together.

box1-4c1-chaincut··v1 4c1fourcardchainseamtreelinkxsim
FREELEAF_V2_20261020 Running

4C1 free leaves v2: depth-1 (host) and every level (RTL)

Depth-1 free leaves: accept 3.9388 -> 4.0850 (+3.71%) on 294 card trees, gate GREEN 28/28; the extra GPU drafter forward costs 0.31 ms

Host mode EAGLE4_4C1_FREELEAF_D1 is built and gated; card A/B pending. RTL every-level form in progress.

box1-4c1-freeleaf··v3 4c1fourcardtreehostfreeleafgpu
ATT_G5KVC_20261020 Running

KV commit + G5FAST in one ATT image: 0xE4B712BB

0xE4B712BB = the 0x12B2 KV-commit ATT + HEAD_PAR 16 (G5 3.00 x ctx + 258), two builds with the kvf + margin recipe

The two ATT features merge with one RTL line; unit, law, md7, elab and host gates pass with RED arms failing; full-drafter replay and timing pending.

box1-4c1-g5kvc··v1 4c1fourcardkvcommitg5fastattbuild
HUMANEVAL_VS_MTBENCH_ACCEPT_20261020 Running

GPU reference accept grid, humaneval vs mt_bench

GPU fa3 grid done on both datasets; card Part B waits for AC#26

mt_bench W2 d7-d15: FPGA accept 4.5-5.6% below GPU fa3; humaneval MAXNEW 32 rows are capped

box1-4c1-gpuref··v1 4c1gpu-refacceptdrafter
ADAPTIVE_TREE_20261020 Done

4C1 adaptive tree, free leaves and hedge: offline replays

Free root leaves in the TREE_STOP phantom slots: accept 3.9388 -> 4.0306 (+2.33%) on 294 card trees, zero card and verify cost; gate GREEN 13/13

Theta and per-level width are flat on mt_bench d7; the drafter's rank-3.. children as free verify leaves add +2.3% (root), +3.7% (depth<=1), +5.0% (every level) accept; hedge h(2,2) >= 0.116 on card trees.

box1-adaptive-tree··v1 4c1fourcardtreehostreplayoffline
VERIFY_KERNELS_20261020 Done

4C1 verify kernels K3/K1/K2/K4 at M<=20

K3 norm bitwise, graph -26 us, texts 50/50; K1 attention graph -352 us but mt_bench texts 49/50 (0.28-logit tie) -> stopped

K3 passes the texts gate. K1 cuts the verify graph 10.8% but flips one mt_bench near-tie (HF fp32 picks K1's token); adoption is the user's ruling. K2/K4 Triton GEMMs lose to cuBLAS.

box1-4c1-vkern··v1 4c1fourcardgpuverifyhosttriton
DUAL_ATT_RTL_20261020 Running

One ATT image for two drafters: Llama-3.2-3B and Qwen3.8-27B by register

Llama mode bit- and cycle-identical to the 0x1288 engine in the attention gate; Qwen 12x128 matches a reference engine; images 0xE4B7129A / 0x129B staged

The 4C1 ATT card now elaborates a max-shape attention (16 slots x 128) and switches between the Llama 16x96 and Qwen 12x128 geometry, rope table, softmax scale, KVI block size, eps, SubLN and e_norm fixes from a MODEL_CFG register block at 0x200. Every modular gate is green; the PDI build waits for a qualifying host.

dual-att agent (box1)··v1 4c1attqwendual-modelattentionropesublnxsim
KVCOMMIT_IMAGE_20261020 Running

KV commit in an image: ATT 0xE4B712B2

0xE4B712B2 = the on-card 0x1288 ATT recipe + KV_COMMIT, building on olegvm and midyard with the margin recipe inside impl_1

Source and stage differ from 0x1288 only in the four KVCOMMIT files; host-side offline gates pass with RED arms failing; timing pending.

box1-4c1-kvcommit-image··v1 4c1fourcardkvcommitbuildatt
WAVEB_DEEP_SWEEP_20261020 Done

4C1 wave B deep sweep d8-d15

w2 d9 V20 525.9 tok/s @ 3.97 (1 round); w2 plateau d8-d10, cards hidden (margin 2.6-2.9 ms); w3 deep slower; stopped on ATT0 item error

Depth 8-12 at W=2 raises accept to 4.06 at flat step time; W=3 deep loses tok/s. Sweep stopped at w3 d10 V27 (ATT0 0x1CC[1] item error); cold AC needed.

box1-4c1-waveB··v1 4c1fourcardsiliconsweepwave-b
WAVEB_SWEEP_20261020 Done

4C1 wave B depth x width sweep

Best cell w2 d7 V20 524.9 tok/s @ 3.86 (4 rounds); W=2 wins at every depth; W=4 d6-d7 leaves 0.3-0.7 ms GPU-wait margin

The optimum stays at w2 d7; W=3 hides the cards (margin >= 1.3 ms), W=4 at d6-d7 nearly exposes them. Sweep stopped at a card fault (MLPV 0x1F4).

box1-4c1-waveB··v1 4c1fourcardsiliconsweepwave-b
MLPV_G11GLM_V2_20261020 Running

4C1 MLPV G11+GLM v2: GLM 256 lanes + SUMSQ fold, 1,658 -> 489 cycles

Full-drafter xsim: G11+GLM 489 cycles (vs 694 for 0x12E1, 1,658 base), bit-exact; 0x12E1 found staged without the lmdown flush fix -> quarantined, corrected 0xE

GLM 256 lanes removes the GLM tail (227 -> 32 cycles); the SUMSQ fold saves 10 cycles. Two images building; no routed timing yet.

box1-g11glm··v1 4c1mlpvg11glmxsim
CHAIN_CENSUS_20261020 Done

4C1 chain census: where the w1 level goes

w1 d7 level 310.4 us: 203.6 counted + 106.8 in uncounted card states/hops; host 0; card join ends with 2.3-2.7 ms GPU wait left

The unattributed level time is on-card (seam path 37.6, G8 23.1, LVL_SEL 15.6, LVL_EMBG 15.5, return hop 8.8 us); at w2 siblings overlap and the level is ATT-bound; D1 is not what hides card cuts, the GPU verify is.

box1-4c1-chain-census··v1 4c1fourcardsiliconcensusloopd1
WAVEB_MEASURE_20261020 Done

4C1 wave B measured: MLPV v1.2

MLPV G9 3,087 -> 850 cycles per item; drafter level 353 -> 316 us (w1 d7); end to end 513.20 tok/s @ 3.86 (7 rounds)

v1.2 cuts the MLPV item from 113 to 72 us and the card loop per level by 7-11%, but the record cell moves only +0.3% to +2.2% (inside one sd).

box1-4c1-waveB··v1 4c1fourcardsiliconmlpvwave-bcensus
I1_VERIFY_CENSUS_20261019 Done

4C1 I1: verify CUDA graph census and VQSTRIDE

Verify graph 3,247 us, 428 kernels, 0 bubbles; VQSTRIDE -152 us graph, exact, card 514.54 vs 511.41 tok/s

The verify graph has no launch bubbles; its time is GEMM (68%), attention (17%) and norms (6.5%). VQSTRIDE drops two dead kernels per layer, bit-exact; default off.

box1-i1-verify-census··v1 4c1fourcardgpuverifyhostab
QWEN27B_DRAFTER_20261020 Running

Qwen3.8-27B + BitNet/SubLN EAGLE3 drafter on the H200: serving and accept

Drafter served bit-exact in SGLang 0.5.16; tree verify works on the DeltaNet target; ours w4d5V24 110.19 tok/s @ 3.65 vs MTP d5 110.97 @ 3.49 vs target 56.73

Our drafter wins accept whenever a tree is allowed (up to 4.22 mt_bench, 7.40 humaneval) but the native MTP head out-accepts it at equal chain depth (d7 3.70 vs 2.85); tok/s at batch 1 is a tie with MTP.

box1-qwen··v1 qwen3.8-27beagle3bitnetsublnsglangh200gpu-onlymtpdrafter-program-target
QWEN27B_HOST_CONVERTER_20261020 Done

Qwen3.8-27B drafter on the 4C1 pair: converter, card-numerics golden, host profile

Regenerated for dual-ATT 0xE4B7129A/B: split-slot weights, rope boot-DMA, KVI gearbox, fixed-e_norm EMBFOLD; 6/6 ranks verify PASS; E2E inputs 18 trees md15

Weights + tables + goldens + host profile for Qwen mode are ready; RTL items: 14-bit scale frac, e_norm wrap on tokens 11/13/15, cheap-head K_C.

box1 qwen27b host/data agent··v2
URAM_HEADROOM_FOR_RETRAINS_20261020 Done

4C1 headroom for bigger drafter retrains

Weight banks use 512 of 1,925 URAM per card; ATT banks 86.7% full, MLPV 62.0%; built ctx window 768; 27B GPU verify 21.13 ms p50

URAM is not the limit for a bigger drafter; every shape change needs a rebuild, and the cost is card time on a serial tree-then-verify chain.

box1-qwen-headroom··v1 4c1fourcardqwen27bretrainuramresources
no new page this day
Mon 2026
ATT_G5_SPEED_20261019 Done

4C1 ATT G5 attention: 16 lockstep head engines halve the per-context cost

G5 cycles 6.00 x ctx + 306 -> 3.00 x ctx + 258, bit-exact; image 0xE4B712E2 closed at true +0.040 / +0.011 (PDI md5 73e7e363).

HEAD_PAR 16 / G5_HPS 1 halves the G5 slope bit-exact; the full image closes once the KVIFAN ch0 300 MHz class is pinned to SLR0.

box1-att-g5fast··v2
D1_HOSTTIME_20261019 Done

4C1 D1 host time: lean attention breaks (T) + cuBLAS pre-warm (D)

T host 6.48 -> 3.71 ms, D host 2.86 -> 1.89 ms; LEANATTN +2.8% mean (5/6 rounds), texts identical 24/24

Both host cuts are exact and measured; after LEANATTN the T phase is GPU-bound (5.0-5.6 ms H200 span), so further T gains need its GPU work split across card windows.

box1-4c1-d1host··v1 4c1fourcardhostsiliconabd1
DW_HOST_PATH_20261019 Running

4C1 direct write: KV rows and root input written straight into the ATT card

Host kit EAGLE4_4C1_DW=1 writes KV rows and the root row into the 0xE4B712D1 s_axi_dw window by QDMA MM; full-drafter xsim with real weights is bit-exact to the

Host path and RTL agree bit-for-bit offline and in xsim; silicon A/B waits for the 0xE4B712D1 flash.

box1-dw-host··v1
RECORD_CELL_20261019 Done

4C1 record cell: every lever on

474.72 vs 459.50 tok/s mean with VERIFY_PRE (+3.3%), 5/6 rounds, texts identical 12/12, accept 3.86

VERIFY_PRE stays on the card-published fast path at theta -2.5 and wins; the 4C1 launcher now defaults the five levers on (plain launch 476.67 @ 3.86).

box1-4c1-recordcell··v1 4c1fourcardhostsiliconabrecord
RTL_REF_WAVEAB_20261019 Running

4C1 exact 2-bit references for wave A and wave B (own-RTL co-sim)

Wave-A silicon trees byte-identical to v0 silicon (6/6 PASS_EXACT vs v0 RTL ref); own-RTL co-sims for wave A/B running, straps 122/127 params exact

Gate is now label-bound: run_4c1.sh maps 0x1288+0x1287 -> wave-A ref, 0x1288+0x128B -> wave-B ref, and refuses (rc 3) a ref built for other images. References land when the co-sims finish.

box1 rtl-ref-waveab agent (4C1 drafter program)··v1 4c1draftergoldenxsimcosimexact
VERIFY_PRE_20261019 Done

4C1 VERIFY_PRE: GPU verify pre-enqueued during the card wait

504.53 vs 453.84 tok/s mean (+11.2%), median +6.3%, 5/5 rounds, outputs byte-identical

The verify's host work runs under the card wait behind a GPU wait on a card-written seq word; step p50 6.350 -> 5.794 ms. Recommended for the record cell.

box1-4c1-verifypre··v1 4c1fourcardhostsiliconab
TREE_STOP_AB_20261019 Running

4C1 TREE_STOP theta: accept versus step time on silicon

theta -2.5 is fastest: 484.75 tok/s at accept 3.86 vs 473.00 at 3.65 for the old -1.25 (8 rounds x 4 arms, identical texts); golden gate not passed at depth >=

Adopt theta -2.5 for the 4C1 path (+2.5% mean, paired +11.75, 5/8 rounds); texts unchanged. Golden depth >= 1 still FAILs identically at -2.5.

box1-research··v1
4C1_MLPV_G11GLM_20261019 Running

MLPV G11 + GLM cut: final norm and low-rank lm-down 1,656 -> 692 cycles, bit-exact

S_G11+S_GLM 1,656 -> 692 cycles per MLPV item (2.39x) in full-drafter xsim, bit-exact; image 0xE4B712E1 building

Three sequencing straps plus 128 GLM lanes cut G11+GLM from 1,656 to 692 cycles per item, bit-exact against the 0x128B RTL in a full-drafter xsim whose base arm reproduces the silicon census (842/814); RED arms bite; OOC lmdown 128 lanes +1.932/+0.015 ns; full-card build running.

box1 G11/GLM agent (4C1)··v2 4c1mlpvg11glmxsimooc
ACCEPT_WALK_20261019 Done

4-card drafter accept walk: card trees vs GPU trees, step by step

TREE_STOP=0 on silicon: accept 3.65 -> 3.94; the card trees match the GPU drafter within 0.006/step once the tree is not stopped

The unexplained 0.35 is the host TREE_STOP lever (85%) and the fa4 lever-Y first drafts (12%); survivor selection, verify pick and KV/RoPE at depth are not the gap.

box1-accept-walk··v1
4C1_HOST_GLUE_CUT_20261019 Done

4-card drafter: cutting the card-process host glue

GLUE_MMIO fast register path: 471.92 vs 441.11 tok/s (+7.0%) @ 3.65, outputs byte-identical 6/6; shared-memory IPC exact but no gain

A one-ctypes-access register path in the card process (same mapping, every guard kept) cut the tree pre-launch 0.306 -> 0.232 ms and KV_LOAD 0.237 -> 0.193 ms per step and raised end-to-end throughput 7.0% mean over 6 rotated rounds with identical texts and per-request meta. Replacing the pickled socket with a memfd ring is exact but did not speed the reply on silicon. The host verify launch (~0.4 ms) is untouched.

box1 host-glue agent (4C1)··v1 4c1fourcardhostsiliconab
BACKEND_KC_AB_20261019 Done

4C1 card A/B: draft backend, first GPU draft and K_C

fa4 + first-GPU-draft + K_C 64 stays fastest: 477.7 tok/s @ 3.65; fa3 449.6 @ 3.67; K_C 256 427.3 @ 3.63

On the 4-card drafter (ATT 0x1288 + MLPV 0x1287), no variant beat the current stack. fa3 raises accept by 0.02 but loses 6% tok/s; turning the first GPU draft off lowers accept to 3.62; K_C 128/256 run on the cards but accept does not rise (3.64/3.63). Measure-only (depth golden unproven).

box1 agent (4C1)··v1 4c1fourcardsiliconabkcbackend
ACCEPT_GAP_SPLIT_20261019 Done

4-card drafter accept gap 3.65 vs 4.04, split on the GPU

All card numerics on the GPU give 4.00 vs 4.04: quantisation is not the gap; 0.35 of the 0.39 is outside weights and formats

2-bit weights cost 0 (the GPU already runs them); card head at K_C 64 -0.08; all card formats -0.04; card silicon 3.65 is 0.35 below its GPU numerics twin.

box1-accept-gap-split··v1
KV_PROMPT_ONLY_20261019 Done

Prompt-only K/V import: card-computed and drafted K/V for accepted tokens

Reusing the drafted K/V of accepted tokens costs 0.05-0.07 accept at mt w2 d7 and 0.00 at humaneval w2 d15 (GPU emulation); card-computed accepted rows give ide

The cards need only the prompt K/V from the GPU. Accept is preserved by both on-card variants; the speed of A is bound by per-item host choreography today.

box1-4c1-kvpromptonly··v1 4c1kvikv-cacheacceptkvcommit
D1_PREFILL_OVERLAP_20261019 Done

4C1 host levers: next-request prefill in the card gaps, deferred KV import, prefix-resident KV

Y_DEFER + D1 on the CUDAPACK baseline: 421.95 -> 461.89 tok/s mean (3 rounds each, accept 3.65, texts and per-request verify counts identical); D1 alone on the

D1 and Y_DEFER are exact and faster; prefix-resident KV with the SGLang radix cache is slower (target prefill +3 ms) and drifts 1 of 50 texts.

box1-4c1-d1··v1 4c1hostprefill
4C1_MORNING_20261019 Done

4-card drafter vs 2-card image: end-to-end, raw drafter and per-phase

4-card 441.8 tok/s vs 2-card 350.16 @ accept 3.65 (+26%); steady step 8.63 -> ~6.5 ms; attention item 1.5x faster, MLP item 1.04x

The 4-card system (ATT 0xE4B71288 x2 + MLPV 0xE4B71287 x2) runs 441.8 tok/s mean (6 runs) vs the 2-card record 350.16 at identical accept 3.65. The draft phase dropped 46%; the attention side is 1.5x faster per item, the MLP/lm-head side only 1.04x (128 lanes @66 MHz vs the 2-card TAIL's 256 @75 MHz). Measure-only: golden at depth vs bf16 unproven; silicon == RTL through tree level L1.

box1 orchestrator (4C1)··v4 4c1fourcardsiliconsweepcensus
G5_CTX_SCALING_20261019 Running

G5 attention context scaling on the 4C1 ATT card

G5 = 6.00 x ctx + 306 cycles; HEAD_PAR 16 gives 3.00 x ctx + 258, bit-exact

Binding = 3 serial passes x 2 head groups. HP16/HPS1 halves the slope; OOC cost +2,360 DSP, +176 URAM. Not built.

fpga-eagle4-accelerator··v1 4c1attg5attentionxsim
RCA_ATT_G4_20261019 Done

ATT G4: silicon 1,074-1,097 vs xsim 1,843 cycles

Silicon ATT runs the embedding fold; xsim ran it off. Fold ON = 1,027 in xsim

The counter is state-gated and correct. Fold ON gives 1,027 zero-wait; silicon's extra 47-70 cycles is consistent with host-row fold-fetch wait (unmeasured on card).

fpga-eagle4-accelerator··v1 4c1attg4embfoldxsim
ONCHIP_PREFILL_WCOL_20261019 Done

On-chip prompt K/V with a W-column engine (4C1 ATT)

K/V-only sweep costs +3.2% at W=8 (xsim); SLICE binds the fit

Cards can compute prompt K/V; a W=8 sweep is 2,399 vs 2,324 array cycles. W=2 fits the 0x1288 ATT; W=4/8 at LANES 256 do not (SLICE).

fpga-eagle4-accelerator··v1 4c1kvsiblingxsim
RESEARCH_SIBLING_PER_PAIR_20261019 Running

One sibling per pair: data-parallel siblings on the 4 cards

No K/V crosses between pairs (lane-chain); the record pair through its host relay ties wave A within noise, and the measured free-sibling bound at w2 d7 is abou

Sibling-per-pair is sound but pays only if the pair round shrinks (hop instead of host relay) or the tree widens; first measurement: wave A w1 vs w2 repeats at d7 and d4.

box1-research··v1 4c1siblingdata-paralleltree-shapeaccept
RESEARCH_HOST_GLUE_FSM_20261019 Running

4C1 host glue into an on-card control FSM

The next tree waits on the 0.381 ms KV_LOAD choreography, not on root-image prep; a card step sequencer plus an in-process driver is worth 0.55-0.70 ms of a 7.0

Start with the KV_LOAD autopilot; GPU-to-card stores need one probe, card-to-GPU push is blocked in the current images.

box1-research··v1
RESEARCH_TREE_SHAPE_MODEL_20261019 Running

4C1 tree shape: which width, depth and verify-token count is fastest

Measured best shape is w2 d7 (378.95 tok/s over 6 wave-A rounds @ 3.65); the model keeps it on top for every card speed and names the w3 silicon ladder and a ve

w2 d7 is the measured best; w3 d7 is the only shape with a path above it and needs two unmeasured parameters; width >= 4 does not pay at depth <= 7.

box1-research··v1
TALON_SPEEDS_0XE4B71262 Done

talon 8x V80 image 0xE4B71262: served-model speeds

0xE4B71262 batch-1 [M]: Llama-3.2-1B 398.8, Llama-3.2-3B 160.5, Llama-3.1-8B 97.4 (TP8); Llama-3B TP6 80.0; Qwen3.8-27B + EAGLE3 16.2, Qwen3.6-27B 12.7 tok/s; D

Batch-1 median of 3 repeats for every model served on talon image 0xE4B71262 (v8, VPU4 on); all exact vs int sim or golden.

hq-research-T2-SW··v1 track2talon0xe4b71262speed
no new page this day
Sun 2026
no new page this day
T2_TALON_TP8_LLAMA_20261018 Done

Llama 1B / 3B / 8B on 8 V80 cards (talon, TP8 ring)

Llama-3.2-3B 155-157 tok/s batch-1, bit-exact, 8x V80

Real Llama models run tensor-parallel on all 8 talon V80 cards over the QSFP ring, bit-exact to the integer simulator.

hq-research-track2··v1 track2talonring8tp8llamasilicon
no new page this day
Sat 2026
BUILD_REPORTS_INDEX Done

Build reports (read-only)

11 report pages

Index of read-only build-report browsers; every file linkable by line.

hq-operations··v17 build-reportsindexread-only
BUILD_REPORTS_1262_V8_S2 Done

Build reports: 1262_v8_s2 (2026-10-17)

5 files

Read-only copies of the 1262_v8_s2 build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_1261_V6_S1B Done

Build reports: 1261_v6_s1b (2026-10-17)

5 files

Read-only copies of the 1261_v6_s1b build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_URE_V2E Done

Build reports: ure_v2e (2026-10-17)

16 files

Read-only copies of the ure_v2e build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_1263_V9B_S1 Done

Build reports: 1263_v9b_s1 (2026-10-17)

3 files

Read-only copies of the 1263_v9b_s1 build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_1262_V8_S1 Done

Build reports: 1262_v8_s1 (2026-10-17)

4 files

Read-only copies of the 1262_v8_s1 build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v2 build-reportstimingread-only
BUILD_REPORTS_125F_KITB Done

Build reports: 125F_kitB (2026-10-17)

122 files

Read-only copies of the 125F_kitB build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v2 build-reportstimingread-only
BUILD_REPORTS_125F_KITBM Done

Build reports: 125F_kitBM (2026-10-17)

121 files

Read-only copies of the 125F_kitBM build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v2 build-reportstimingread-only
BUILD_REPORTS_COUNCIL_2026_10_17 Done

Council results (2026-10-17)

8 files

Read-only copies of the council-2026-10-17 build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v5 build-reportstimingread-only
TRACK1_4C1_FLOORPLAN_20261017 Running

Track 1: 4C1 engine floorplan per card (failed timing builds)

Clickable engine floorplans of both failed 4C1 checkpoints: link layer fails on 402.83 MHz on both cards; MLPV tree engine at +0.003 ns is the next gate risk

ATT and MLPV failed only on the link clock; MLPV u_tree_level +0.003 ns misses the +0.03 release gate; ATT GEMV array +0.234 ns is its tightest drafter path.

box1-main··v2
BUILD_REPORTS_125F_S1C Done

Build reports: 125F_s1c (2026-10-17)

129 files

Read-only copies of the 125F_s1c build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_1256_S1R Done

Build reports: 1256_s1r (2026-10-17)

127 files

Read-only copies of the 1256_s1r build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_1254_S2_GOOD Done

Build reports: 1254_s2_GOOD (2026-10-17)

72 files

Read-only copies of the 1254_s2_GOOD build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v2 build-reportstimingread-only
TRACK1_4C1_ATT_S3LAATTF2 Running

Track 1: 4C1-ATT per-phase architecture, read from the failed timing build only

4C1-ATT build s3LAattf2 (0xE4B71271): 256 lanes x 1 dual-port URAM each = 32 weights/lane/array-cycle; G4 floor 1,728 and G6 288 drafter cycles per child; K cac

Every number on this page comes from build s3LAattf2: its staged RTL, its synth log, its placed utilization and its post-placement checkpoint. The GEMV phases are URAM-read bound by construction; the build failed timing only on the link clock (fixed in the relaunch).

box1-main··v1
RING8_TIMING_DIAG_20261017 Done

Ring-8 Timing Diagnosis

v8 0xE4B71262: 52/52 x8 cards, ring BIST PASS [M]

v8 0xE4B71262 (C_COLS 2 @225, GT island, crumbc fanout) closed timing (+0.003 / true +0.093, 0 failing) and passes on silicon: HBMS ladder 52/52 on all 8 talon cards, ring BIST 16/16 PASS 0 errors

hq-research-track2··v6 track2talonring8timinghbms
no new page this day
Fri 2026
TRACK1_4C1_ARCH_20261016 Running

Track 1: the 4-card drafter architecture shift (record pair to 4C1)

Record pair -> 4C1 (attention TP2 pair + MLP/vocab TP2 pair over QSFP); splits bit-exact in xsim at ~0.5x cycles per card; ATT 0xE4B71271 building on box1, MLPV

The 4C1 shift moves G7 off the HEAD, halves every weight-bound phase per card with 48-bit partial reduces over QSFP, and replaces the host relay with a 6 KB card-to-card seam; SLR1 (all QSFP GTMs) is the die to keep below 80%.

box1-main··v2
FOURCARD_DESIGN_20261016 Running

4-card drafter: where the time goes and how to slice the weights across 4 V80s

Attention TP2, MLP TP2 and vocab shard (replay merge) are bit-exact in xsim; TP2 halves body cycles per card; the vocab shard saves ~11.7K cycles per child at K

Record pair: cards 28% of the step; HEAD GEMV-bound (G4+G7 78%), TAIL TREE+G9-bound (81%); 4-card slicing P1-P4 in simulation.

box1-main··v3
no new page this day
no new page this day
Thu 2026
ARCH_BOOK Running

EAGLE4 Architecture Book

66 images across 3 tracks, one interactive block diagram per image

Pick a track and an image to see its block diagram, what it added, its timing and its measured tok/s.

hq-operations··v5 architecturepinneduitrack1track2track3
TALON_SERVE_20261015 Running

Serving Llama on talon V80s with HBM streaming: 1B to 70B, no GPU

Spec decode 8B+1B 51.1 tok/s and 3B+1B 81.9 tok/s on V80s; 70B on 8 V80s 2.34 [M]

The flashed HBMS image 0xE4B71254 serves real Llama models end to end on silicon with no new PDI: one card up to 8B, eight cards pipeline-parallel for 70B; decode is bound by W8 weight streaming at ~230 GB/s per card.

hq-operations··v6 track2talonhbmsservingllamasilicon
TALON_8CARD_SERVING_20261015 Running

Talon 8-card serving: one model across 8 V80s

8 x Llama-3.2-3B: 439 tok/s batch 8 / aggregate [M]; batch-1 across 8 cards needs 0xE4B71256

Every talon card serves 3B bit-exact (54.8-57 tok/s each); one model across 8 cards waits on the HBMS all-reduce to ring bridge (0xE4B71256).

hq-research··v1 track2talonring8servingmulticardarchitecture
TALON_HBMS_ARCH_20261015 Running

Talon V80 HBMS engine: architecture, Vivado results and silicon cycle profile

Llama-3.2-3B 54.8 tok/s batch-1 on one V80; 94.2% of cycles in W8 weight streaming [M]

Batch-1 decode on the 0x1254 s2 engine is bound by the 32-port weight-stream beat rate (230 GB/s), not HBM (stall 3.7-3.9%) or DSPs (3 of 4 columns idle).

hq-research··v2 track2talonv80hbmsarchitecturecyclesvivadollama
no new page this day
Wed 2026
SPEC-DMX-20261014 Done

Speculative decoding: our EAGLE4 FPGA drafter vs d-Matrix Corsair

Both split draft (accelerator) from verify (GPU). d-Matrix runs a separate 1.6B draft LM as a token chain from on-chip SRAM; we run EAGLE4, a 1-layer feature-le

Same placement idea, different drafting algorithm, model-size ratio and memory system. Their claim (2-10x vs GPU spec decode, measured+modeled) and our batch-1 tok/s are not comparable numbers.

box1-main··v1 spec-decodeeagle4dmatrixcomparisonprogram-1
no new page this day
TUNIT_0x1258 Done

Constant-fold ternary unit on one V80: TriLM-99M layer 0, bit-exact on silicon

0xE4B71258 TUNIT: 7 ternary projections (3.01 M weights) folded into LUTRAM, 0 mismatches / 10,240 on silicon, 266 cycles per projection [M]

The single-image LUTRAM fold is correct on silicon and its compute time matches simulation; it is one layer's projections, so it cannot generate text by itself.

hq-operations··v1 ternaryconstant-foldtrack-tsilicon
Sun 2026
DRAFTER_INT8ACT_20261011 Running

8-bit activations in the Llama-3.2-3B EAGLE4 drafter (Program 1)

int8 activations cost no accept: baseline 8.28 / 8.29 vs int8 body 8.26, +fc+lm 8.29, group-128 8.31; RED int2 1.40 (humaneval 80, d15 w4, fa4)

The drafter's ternary body linears accept BitNet-style int8 activations with no accept loss on the pinned GPU command, so its GEMVs can use the integer (int8 x ternary) contract that constant folding (Track T) assumes.

box1-golden··v4 eagle4drafterint8activationternarysglang
GOLDEN_SUITE_20261011 Running

Golden captures and RTL verification suites (BitNet-2B-4T, Qwen3.6-27B, EAGLE4 drafter)

Capture bug (non-causal prefill) found + fixed; v3 H200 goldens GREEN on CUDA (DeltaNet 48/48, gated GQA 16/16, BitLinear 210/210, every RED caught); run_tier_q

Running. Causal-fixed v3 goldens for both target models pass every Tier-Q check on the capture device with REDs caught; four numeric contracts handed to HQ and adopted; Tier-E hook ready for the first RING8 builds.

box1-golden··v5 goldenverificationtestbenchbitnetqwenaxiospreytier-e
OSPREY_GOLDENS_20261011 Superseded

OSPREY per-op goldens + 3STAGE seam vectors (Program 1)

Tier-Q 493/493 checks: GREEN passes and every planted RED is caught (0 vacuous); F1 trees match the FA3 GPU trees on 24/24 levels up to near-ties

Per-op goldens (every phase G1..G8 + TREE) at ctx 16..564, md 1/2/3, for 4+4 prompts, plus the A->B / B->C / C->A seam regions as .bin/.mem; the checker catches every planted defect.

box1-golden-osprey··v2 goldeneagle4drafterosprey3stage
AXI_TB_20261011 Superseded

Golden AXI suite: protocol checkers, BFMs, AXI input testbench

Superseded by box1-golden-suite-20261011 | Suite self-test 20/20 PASS (3 GREEN seeds x 84 checks LIVE, 17/17 RED arms detected); control_regs strict B_DEP/R_DEP

Reusable AXI4/AXI-Lite/AXIS checkers + randomized BFMs; every check proven to fire. control_regs.sv passes all 60 ABI vectors but answers B/R in the same cycle as the request handshake (A3.3.1 strict dependency FAIL).

box1-golden··v2 axitestbenchverification
RING8_20261011 Running

8-FPGA HBM ring LLM serving with a model contract

Two 8-card image tracks: T (ternary fold) and Q (Qwen3.6-27B HBM flux); [M] QSFP 99.84 Gb/s per direction, 1.20 us echo; binding item = R8F re-floorplan (0x1235

Running. Umbrella page of the 8-card FULL TARGET program: Track T -> ternary-make-possible-20261011, Track Q -> qwen27b-ring8-20261011. HQ build plan v2 sets the gates and the L0-L7 ladder; next: L0 HBM readout on card 61, R8F re-floorplan, ring cabling card i P4 -> i+1 P3.

box1-golden··v3 ring8hbmcontracttppp
QWEN27B_RING8_20261011 Running

Track Q: Qwen3.6-27B on 8 V80s (HBM-maximized flux, QSFP ring + PCIe FPGA-to-FPGA DMA)

TP8 schema checked against the real Qwen3.6-27B shapes; stock shell can RECEIVE P2P but has no SEND route, CIPS probe [M] names the fix (NOC_CPM_PCIE_0); 0x1250

Running (reopened as the Track Q living page). Shape, schema and shell facts measured; HQ e8 transport xsim-exact under faults. Next: R8F re-floorplan + lanes-vs-HBM winner -> 0x1250 C4 build on the laptop.

box1-golden··v3 qwenring8hbmp2ppcieqsfptp8track-qmtp
QWEN_HBM_FLUX_8CARD_20261011 Superseded

Track Q: Qwen3.6-27B on 8 FPGAs — HBM-maximized flux, QSFP ring + PCIe FPGA-to-FPGA DMA

Superseded by qwen27b-ring8-20261011 | TP8 schema verified against the real Qwen3.6-27B shapes; the stock V80 shell has the P2P RX path but no TX path — CPM5 sl

Program 2: 8-card FULL TARGET; Track Q. Design + shell facts measured; no image built yet (Qwen PDI builds owned by the laptop copy of box1-golden, after the 3STAGE report).

box1-golden··v2 ring8qwenhbmp2ppcieqsfptp8
TERNARY_MAKE_POSSIBLE_20261011 Running

Ternary constant folding on 8 V80s: measured levers, full reproduction pack (Track T)

[M] 4-layer TriLM-99M card CLOSES 400 MHz OOC (WNS +0.021, 0 failing; SLR2 empty) → 4 cards instead of 8. Unit PDI 0xE4B71258 PASSES on silicon (10240/10240).

Running. Silicon: 0xE4B71258 bit-exact on V80 card 42. OOC: 4 single-image layers per card close 400 MHz in two SLRs, leaving SLR2 free, so TriLM-99M fits 4 cards. Next: shell integration, on-chip activation path, ring.

box1-golden-laptop··v17 ternarybitnettrilmconstfoldmulticontextreproductiontrack-t
TERNARY_CONSTFOLD_20261011 Superseded

Ternary constant-folded LUT adder trees (BitNet/TriLM)

Superseded by ternary-make-possible-20261011 | Full target on 8 cards: a folded ternary neuron is bit-exact on real BitNet-2B4T weights; bit-serial costs 1.19 L

Works bit-exact; the wall is area (~1.2 LUT per nonzero bit-serial). Road to a PDI: folded unit PDI, one-layer-per-card routability, sparsity-trained tiny folds; URAM-resident path for 2B4T.

box1-golden··v4 ternarybitnetconstfoldlut
TERNARY_FOLD_8CARD_20261011 Superseded

Track T: constant folding on 8 FPGAs — TriLM-99M fully folded + a co-resident folded drafter

Superseded by ternary-make-possible-20261011 | Multi-context LUT6-INIT folding = 0.0796 LUT/weight (10.4× denser than single-context), so TriLM-99M (48.2 M tern

Program 2: 8-card FULL TARGET; Track T. Re-aimed per user: (1) TriLM-99M fully folded on 8 cards, (2) a constant-folded ternary drafter co-resident with the HBM target. BitNet-2B-4T stays the URAM-resident ternary reference.

box1-golden··v3 ternarybitnetconstfoldring8track-t
Wed 2026
ARCH_4CARD_VARIANTS_20261007 Done

EAGLE4 on Four V80s: Variants and Choices

SIBX-E 4-card rejected (X 114.35 vs P 163.47 tok/s [M]); CL4 ring chosen, untested; record still 339.12 [M] vs GPU-only 446-486 [M]

Diff from the 10-05 architecture post: what each 4-card variant became, the measured verdicts, and why CL4 ring + shape-generic + HBM streaming are the next bets.

hq-junior··v1 weeklyarchitecturefour-cardv80decisions
no new page this day
no new page this day
Tue 2026
EAGLE4_TRACE_20261006 Done

EAGLE4 Trace: per-block timelines, bubbles and utilisation

Silicon step: HEAD card busy 5-17 %, top bubble = waiting on GPU [D]

Perfetto traces per hardware block from xsim + box1 hostprof; the draft step is a serial host->cards->GPU chain, so overlap beats engine speed-up.

hq-eagle4-trace··v2 traceperfettoutilizationbubblestooling
no new page this day
no new page this day
Mon 2026
ARCH_EVOLUTION_20261005 Done

EAGLE4 Architecture Evolution

153.4 → 339.12 tok/s [M] in 17 architectural steps

Animated walk-through of the EAGLE4 V80 drafter: HEAD/TAIL split, 256-lane DSP58 GEMV, KV import, LEVEL_PIPE/FG8, 75 MHz TAIL record 339.12 tok/s [M], QSFP D2/CL4 4-card paths [P], and runtime shapes.

arch-audit··v2 architectureanimationweekly
ROADMAP_REVIEW_20261005 Done

EAGLE4 Roadmap Review Pack

Reading list for the roadmap session: 6 themes, site pages + lab files, with discussion questions

Curated pages and files to review with partner hardware leads, grouped by the roadmap decision each one feeds.

hq-junior··v1 weeklyroadmapreview
WEEKLY_ARCH_20261005 Done

EAGLE4 V80 Architecture Week 40

Record 339.12 tok/s [M] batch-1 (HX1 BD + TCLK75); 17 architecture milestones, widths from built wrappers

Weekly architecture reference: current HEAD/TAIL datapaths, memories, clocks and links, with a selectable timeline of every major shift.

hq-junior··v1 weeklyarchitectureauditv80
no new page this day
no new page this day
Wed 2026
PRODUCT_PAGE_20260930 Done

Eagle Companion Chip

1.33x vanilla H200 on HumanEval, 331 tok/s MT-Bench (measured)

On-chip speculative-decoding drafter: 1.15x (MT-Bench) / 1.33x (HumanEval) vanilla H200 at batch 1; FPGA evaluation system, SGLang, switchable target model

hq-junior··v6 productcommercial
no new page this day
no new page this day
Tue 2026
WEIGHT_RESIDENCY_20260929 Done

EAGLE4 Weight Residency

41.58 MiB on-chip; 4 cards ~5x

41.58 MiB 2-bit body on-chip (URAM), fp16 lm-head and embeddings in HBM; 4 V80s hold ~0.9B 2-bit weights, ~5x today

hq-junior··v1 weightsresidencyuramhbm
PROGRESS_20260927_29 Running

EAGLE4 progress 09-27 → 09-29: new batch-1 record 323.30 tok/s

Batch-1 record 323.30 tok/s (HX1+FSL + SPEC0, mt_bench w2 d3) [measured], +5.1% vs 307.67; humaneval w1 d4 382.53 [measured]

A new HEAD (HX1, E3b packed-DSP lanes) and TAIL (FSL, LUTRAM seams) took batch-1 to 323.30 tok/s, exact. The 307.67 miss was a slow-boot host state, now classified in <1 s. SLR1 is the wall; EVAC, FSL and I1 free it.

hq-main (hq-junior)··v1 eagle4progresspairheadtailsiliconresourcesboot
no new page this day
no new page this day
Sun 2026
WEEK_REVIEW_20260924_27 Running

One week on a two-FPGA speculative drafter: what was asked, what got built, and what the silicon said

Batch-1 best 307.67 tok/s (A5+FA+SPEC0, mt_bench w2 d3) [measured], up from 175.9 on 09-24; GPU-only EAGLE reference 485.79 [measured]

Host-side fixes and two image generations took the pair from 175.9 to 307.67 tok/s batch-1 (+75%). Depth 3 now beats depth 2 on both benches. The remaining gap to the GPU (~63% of it) is per-level FPGA cost and the strictly serial GPU/FPGA step.

hq-main (hq-junior)··v8 eagle4reviewpairheadtailsiblingsdeptheducation
JOURNEY_20260924_27 Done

Four days on the EAGLE4 FPGA drafter pair: what was slow, why, and what moved

Best pair A5+F4 268.01 tok/s (1 round), 264.29 (3 rounds) on mt_bench w2 d2 [measured]; GPU-only EAGLE ~447 tok/s

Priming, not depth, was the accept problem; KV import fixed accept at 1.5-1.7x speed. Then host glue, burst seam export and pipelined attention took the pair from 188 to 268 tok/s. What remains is per-level cost (~1.80 ms/level vs <1.72 needed for d3 > d2) and the GPU verify.

hq-main··v2 eagle4journeypairheadtailkvideptharchitecture
no new page this day
no new page this day
Sat 2026
BR055 Done

Tested Images 2026-09-26

Every pair image tested on silicon on 2026-09-26

Best 1-round pair: A5+F4 KVI+FASTGLUE4 mt w2 d2 268.01 tok/s @2.55 [measured]; 6 pairs catalogued with w x d sweeps.

hq-main··v1 catalogsiliconsweep
no new page this day
no new page this day
Thu 2026
SGLANG_DRAFT_KV_LIFECYCLE_20260925 Done

SGLang draft-KV lifecycle vs the pair

Unprimed saturation = no prompt KV; primed w3 predicted d5 3.76 / d15 4.71

SGLang never reuses tree KV; the pair needs priming (cheap via PRIME_BURST) plus a host slot fix; RTL child RoPE is +1 (accept-neutral).

fpga-eagle4-accelerator··v1 acceptkvsglangdepthemulator
HOST_GLUE_STRUCTURAL_20260925 Running

Host glue G: ranked terms and two thread-free / pinned-CPU levers

mock 70/70; card-worker GIL hand-offs 0.58 ms/step vs 0.09 ms piped

Two default-off host levers (RPC_PIPE, CPU_PIN) are byte-identical in mock with 4 REDs caught; they remove host scheduling only, so G < 3 ms still needs SGLang verify work. Silicon A/B pending.

host-overhead-agent··v1 hostgluesglangpairgil
HEAD_G7SEAM_PIPE_20260924 Running

G7 seam write pipeline for a faster HEAD clock

60% of the HEAD's 100 MHz failures are one write-index register fanning out to 512 seam rows; banked write fix passes unit + elab gates

SEAM_WR_PIPE removes the r4 seam write fanout (unit A/B PASS, RED FAILs, scope census confirms placement of the arm). The HEAD clock still cannot rise until attention is pipelined (67.3 MHz ceiling); image 0xE4B711E5 (CD + fix, 66 MHz) and the full-HEAD gate are in flight.

g7seam track agent (EAGLE4 orchestrator fork)··v2 headtimingfmaxseamxsimimage
PAIR_FRSPEC_PREP_20260924 Done

FR-Spec vocab shortlist on the HEAD/TAIL pair: silicon result

No speed gain today: FR-Spec shortens the TAIL walk but the HEAD is the bottleneck; keep K_C 64

On silicon at steady load, the vocab shortlist gives no gain at width 2 (156-157 tok/s at every tier vs 156.44 full) and at most +3% at width 3 (noise level), at a cost of 2-4 extra verify steps out of ~875 on mt_bench and -2.5% accept on humaneval. FR-Spec shortens the TAIL's TREE walk, but the HEAD is the busier card. Re-test once a faster HEAD (CD/E1/E3) is on silicon. K_C: keep 64.

frspec-prep worker (from PAIR_FRSPEC_PREP_20260924.md)··v2 frspecpairvocabtreelm-head
HEAD_PRIME_BATCH_20260924 Running

HEAD PRIME_BURST: many KV-only prime rows per doorbell

Design decided + RTL + host path written; 4 full-HEAD xsim gate arms running on the mesh

Priming costs about 1 ms per row on the pair and the HEAD computes for only about 0.23 ms of that; the rest is the host round trip per row. PRIME_BURST runs N rows from one doorbell, with the card loading each next row itself. The xsim verdict is pending.

head-prime-batch fork··v1 headkvonlyprimexsimrtl
ENGINE_SPLIT_IMAGES_20260924 Running

Engine-split design D: first GEMV-card and state-card images (fit/timing probes)

GEMV card 0xE4B711D0 (role 3, LANES 256) + state card 0xE4B711D1 (role 4, TISEQ) building on olegvm2; both elaborate clean, ESIMG_STAGE PASS

Launched. The routed builds will answer (1) whether the 256-lane GEMV card still fits with the hop link in the netlist and (2) the open question: does the state card's SLR1 hold attention + the link + the Aurora core. Not serving images.

engine-split image fork (from reports/results/ENGINE_SPLIT_IMAGE··v1 engine-splitdesign-dpdi-buildfit-probehop-link
RESEARCH_OVERVIEW_20260924 Running

EAGLE4 research overview — 2026-09-24

19 pages refreshed/published today across 5 tracks: GEMV engine, HEAD levers, pair silicon + accept, engine split, infra

No single track is 'the' bottleneck any more: the array clock (CD, routed), the HEAD seam/KV-only levers (gated, committed), and the pair's tok/s (112.73 @2.22, +10.4%) all moved today. The two open threads that gate the next silicon step are the width-3 accept shortfall (regime, not proven defect) and the 100 MHz fmax census (G7 seams, not the array, are 60% of the failures).

research-overview agent··v1 overviewindex2026-09-24
WHOLE_DRAFTER_FMAX_CENSUS_20260923 Running

What stops the whole drafter clock at 100 and 133 MHz (routed HEAD 0xE4B711C9, TAIL 0xE4B711CA) — HEAD census landed

The 100 MHz HEAD census landed: 542,754 failing endpoints, WNS -4.850 ns. 60% of the failures are the two G7 FF seams (wide register files), not the array or at

The HEAD census at 100 MHz is complete and bucketed by block. The G7-up and G7-gate FF activation seams dominate the failing-endpoint count (218,700 + 104,720 = 60% of 542,754), ahead of attention (66,480, but the single worst path at -4.850 ns) and the GEMV array (32,201, -3.878 ns). A faster whole-drafter clock is therefore a seam fan-out + attention-combine project, not primarily an array project — consistent with the array's own 132/133 MHz domain (image CD) already closing cleanly on its ow

fmax-census agent··v2 timingfmaxroutedcensusclockheadtail
ARRAY_DUALCLK_20260923 Running

2-bit GEMV array on its own 2x clock (132 MHz array, 66 MHz drafter) — CD ROUTED

Full HEAD image 0xE4B711CD is ROUTED: WNS +0.005 / WHS +0.010 ns on BOTH the 66 MHz drafter clock and the 132 MHz array clock. Bit-exact incl. full-HEAD 29/29;

Running the array alone at exactly 2x the drafter clock, fed by a per-die BRAM activation cache, is bit-exact and cuts the HEAD item from 33,010 to 18,610 clk in the real drafter. What was pending at the last update — whether the full HEAD image closes timing in context — is now answered: CD is ROUTED, closing both clocks with positive margin, and even freed some resources (LUT 998,363 vs C8's 1,039,733). Silicon flash and the resulting pair tok/s are still pending the user.

fpga-eagle4-accelerator (array_dualclk agent)··v3 gemvarrayclock-domaincdcxsimtiminghead
SIBLING_W_ON_GEMV_CARD_20260923 Running

Sibling batching W on the design-D GEMV card: one weight sweep for W children

W=4 array placed with 821/516 SLL (August W=8 needed 23,907); W=8 bit-exact in xsim; W=8 at 128 lanes is now blocked by AREA, not wire. W=8 STILL NOT PLACED (ta

Per-die activation replicas + a W-wide serialised return remove the August inter-die-wire wall for the array (measured at W=4, 821 SLL). The new wall is die area: 2,038 LUT per lane-column measured, so W=8 x 128 lanes is ~157% of the routable LUT budget (derived). Once the GEMV sweep is batched, the state card's per-child G12/TREE/G8 tail dominates a level. As of 2026-09-24 ~16:50 EDT the user-mandated target W=8 has NOT been placed at any lane count: the W=8 L64 flow was just admitted on Oleg's

sibling-batching agent (fpga-eagle4-accelerator)··v2 siblingbatchinggemvsllxsimplacementengine-split
ENGINE_SPLIT_STATE_STEP2_20260923 Done

Engine split, design D steps 1–2: GEMV card (role 3) + state card (role 4) + hop_link, bit-exact in xsim

State card bit-exact vs single card: 7,414 items / 0 mismatches; GEMV card 24/24; hop_link 40,800 hops exact

Both halves of the engine split reproduce the single-card drafter bit-for-bit in xsim (md=2, ctx 16, OSPREY 32×96), RED arms fire, and the link transport delivers every hop exactly under back-pressure. No synthesis, timing or silicon yet; the pair speed-up is still a model. Nothing here has regressed as of 2026-09-24 ~16:50 EDT: the separate step-3 joint two-card sim (other agent, fork engine_split_pair_20260923, rebased on this fork's f8f62fb / the OUT-DMA latch fix) shows its old un-rebased co

engine_split_state agent··v2 engine-splitdesign-dstate-cardgemv-cardhop-linkxsimbit-exact
ARRAY_SYSTOLIC_LUTGEMM_20260923 Running

Systolic activation flow and multiply-free LUT-GEMM on the 2-bit GEMV array (module-only)

A2+B L128 route HARVESTED: 171 MHz fmax, keeps A2's clock while cutting u_array LUT −39.6%; L256 systolic-vs-broadcast synth landed, impl queued

Adopt the LUT-GEMM (g=2) lane: exact, cycle-neutral, −138K LUT at LANES 128, and now confirmed on the routed netlist — 171 MHz fmax vs 174 (A2 alone) / 177 (broadcast base), so the LUT saving costs no clock that matters at the 66 MHz ship clock. Do not adopt the systolic activation chain for clock at LANES 128. L256 systolic-vs-broadcast: sys8 synth landed, its impl and the base L256 pair are still in the mesh queue.

array-systolic agent (ARRAY_SYSTOLIC_LUTGEMM_20260923)··v2 gemvarraysystoliclut-gemmt-macxsimtimingsll
CONSOL-20260924 Running

Consolidated next images: HEAD E3 + TAIL 0xE4B711F1

Only gate-GREEN changes: HEAD E3 (fast array clock + wide seam + KV-only skip-attn + pipelined attention) and TAIL CF + T1

Next pair = HEAD E3 + TAIL 0xE4B711F1 (CF + T1), gate-green changes only; building

EAGLE4 consolidation fork··v1 pairimageconsolidationtailhead
PAIR_HOST_OVERHEAD_20260923 + HOST_OVERLAP_IMPL_20260924 Running

Pair host round-trip overhead: waterfall, HOSTFAST, and exact draft/verify overlap

HOSTFAST=1 (+4.8%) still the launcher default. NEW: I7/I8 CUDA-event profiling measured G = 8.58 ms median (default); four new byte-exact cuts take it to 6.28/6

The round-1 HOSTFAST=1 result (+4.8%, 50/50 identical) stands and is the launcher default. Two later missions (I7, I8) added a CUDA-event timing instrument that decomposes every step into V (target-verify GPU), D (tree+KV span) and G (host glue), then measured G directly instead of estimating it: G = 8.58 ms median / 6.00 p10 on the default path. Four new byte-exact cuts (root-image + prep-row CUDA graphs, a cost-fallback fast path, a single-transfer capture) take G to 6.28/5.03, and with the ex

host-overhead agent (a8de43581730ce30f)··v2 hostsglangpairoverlapprofilingqdma
PAIR_PER_OP_XVAL_20260924 Running

Per-op GPU-vs-pair drafter cross-validation tool

Tool built and self-tested (0 FAIL of 17 clean items, 4/4 planted defects correctly localised); silicon run against the real HEAD/TAIL pair is PENDING

The comparator is ready and its own self-test is green, but no card data has been captured yet — whether the pair's depth-3 accept gap (1.97->2.08 vs GPU's 2.64->3.19) is an op-level defect or the unprimed regime is still an open question until the pending silicon capture runs.

per-op-xval agent··v1 cross-validationaccept-gaptoolingselftestpending
HEAD_SEAM256_KVSKIP_20260924 Done

HEAD levers: 256-bit seam export and KV-only attention skip

ALL GATES GREEN, RTL committed (verilog 0220bf7, parent 2a6ea01a): seam export/child 26,434 -> 10,754 (256 b) -> 2,501 cycles (256 b + page bursts); KV-only G5

Both levers are byte-exact in every gate, including all 10 full-HEAD arms (29/29 items, 928/928 KV rows, every RED fails exactly as predicted). RTL is committed (verilog 0220bf7, parent 2a6ea01a). Mode 1 (256-bit port alone) removes ~44% of the silicon export time because 1,216 per-cell B round trips dominate; page-capped bursts (mode 2) reach the model's ~40 us, but sit in R2 PIPE_EXP_MULTI's risk class (the 0x1190 accept-1.00 confound, also byte-identical in xsim), so mode 1 is the first image

fpga-eagle4-accelerator (head-levers agent)··v3 headaxiseamattentionkv-cachexsimpair
ENGINE_SPLIT_PAIR_STEP3_20260923 Running

Engine split design D, step 3: GEMV card + state card over the real hop link, in one xsim

GATE GREEN + committed (verilog f29beae): pair bit-exact vs single card on 24/24 items, md=2 tree 1,356,425 vs 1,540,897 xsim cycles (-12%); 2/3 planted-defect

The two-card protocol works end to end (24 items, 72-node tree, 0 CRC/seq/overflow) and the full-data bit-exact gate is GREEN at L=0 and L=40 (RTL committed verilog f29beae). Hop 3 must carry GLM + G11 (211 wire cycles). The GEMV item FIFO must hold K_F=8. RED arms red_flip and red_tag were harvested 16:10 EDT and both FAIL exactly as designed (non-vacuity proven); red_crd, red_crd1 and the fifo4 deadlock proof are still running on the laptop (slow host, in GEMV weight init since 2026-09-23 21:3

engine-split-pair agent (fpga-eagle4-accelerator)··v3 engine-splitdesign-dxsimhop_linktwo-cardqsfp
GPU_WIDTH_DEPTH_ACCEPT_GRID_20260924 Done

GPU-only accept and tok/s grid: tree width × depth on the pair's checkpoint (H200)

Width 3 keeps 95.5% (mt_bench) / 97.8% (humaneval) of width 8's accept at depth 2; GPU bs2 = 453 tok/s mt d2 w8

Narrow trees lose less accept than the pair model assumed (width 3 keeps 95.5-97.8% at depth 2, falling only to 90.7-93.4% at depth 15). Grid is COMPLETE: 132/132 runs (fa3 40/40, fa4 40/40, anchors, repeats, controls, bs2, E3 split, contamination re-runs). d2 w3 remains the pair's modelled optimum in every scenario and fa4-retention re-run, with a larger margin (+37% vs the model's original +27-35%). GPU bs2 raises the bar ~1.3-1.5x at unchanged accept; batch-2 verify only costs +6-13% over bat

gpu-accept-grid agent··v4 gpuaccepttree-widthdepthsglangh200baseline
PAIR_ACCEPT_GAP_PRIMING_20260924 Running

The pair-vs-GPU accept gap is a priming/regime effect, not a drafter defect

Test B: per-level golden rules out an RTL defect. Test A: primed pair closes to GPU at width 3 (d2 2.61 vs 2.64, d3 3.15 vs 3.19) — the gap is the unprimed KV r

The pair's drafter datapath is bit/rank-correct against the GPU golden (Test B). The accept shortfall vs GPU (2.22 vs 2.78 baseline; 1.97/2.08 vs 2.64/3.19 at width 3, depth 2/3) is explained (92% at width 8, and closed almost entirely at width 3 by Test A) by one mechanism: SGLang's GPU drafter always attends a fully-primed prompt KV, while the pair's default catch-up-only regime starts every request with an EMPTY drafter KV. Cheap KV priming is the lever, not an RTL fix; its own silicon A/B (c

pair-accept-gap agent··v1 acceptprimingkv-cachesglangpairgpu-comparison
TREE_WIDTH_PAIR_20260924 Running

Runtime tree width on the HEAD/TAIL pair: TAIL image 0xE4B711CF for width 3 — ON THE CARD, silicon shortfall found

TAIL CF is flashed and running on silicon (widths 8/4/3/2/1 all GREEN, FRONTIER GATE PASS). Width 3 is +32% tok/s (146.8 vs 111.1 tok/s) but accept only 1.97 (0

Width 3 works end to end on silicon: same HEAD, new TAIL 0xE4B711CF, one AXI-lite register. The pair gains +32% tok/s at width 3 (111.1 -> 146.8 tok/s, mt_bench cap-32, unprimed) even though accept drops more than the GPU-only comparator predicts (1.97 vs a 2.10 GPU-shaped expectation). The extra 6% of accept is explained, not measured, by the pair's known unprimed-regime weakness (it loses proportionally more when only the top-3 of 8 ranks survive). W=2 and W=4 measured too (1.58 and 2.07); the

tree-width agent (fpga-eagle4-accelerator)··v3 treewidthtailxsimsiliconhostpdisglang
PAIR_C8CF_BASELINE_20260924 Done

Pair silicon baseline: HEAD 0xE4B711C8 + TAIL 0xE4B711CF — matched benchmark, width and depth sweeps

112.73 tok/s @ accept 2.22 (was 102.13 @ 2.22) — width 2 peaks tok/s (153.4), width 3 is the tok/s×accept sweet spot; 92% of accept gap is the UNPRIMED regime,

HEAD 0xE4B711C8 (HEAD_PAR 8) + TAIL 0xE4B711CF (TISEQ + TREE_P 16 + runtime width) raised the matched pair benchmark from 102.13 to 112.73 tok/s at unchanged accept 2.22 (+10.4%), by making the HEAD's attention phase ~5.7x faster and the TAIL's tree walk ~1.47x faster on silicon. Width and depth are now characterized end to end: tok/s peaks at width 2 (153.4 @ 1.83), width 3 is close behind (146.8 @ 1.97, +32% tok/s over width 8 for -11% accept) and is the current default candidate; depth costs

pair-silicon-c8cf agent··v1 pairsiliconc8cftree-widthdepthacceptunprimed
G4_EMBFOLD_INTEGRATION_20260924 Running

G4 embedding constant fold: HBM P-table, fetch engine, host path

Fetch engine COMMITTED (verilog f792e27): unit gate run4 all-PASS incl. RED, OOC synth 2,416 LUT / 1,518 FF, WNS +12.25 ns @66 MHz; full-drafter HBM gate still

Pieces 1 (table builder) and 3 (host mock) remain GATED/committed as before. Piece 2 (HBM fetch RTL) is now COMMITTED (verilog f792e27) with a clean run4 unit gate (latency sweep, parallel abort-drain, RED fired) and a first OOC synthesis census. The full-drafter xsim gate that would prove the fold end-to-end has NOT landed yet: 4 arms have been queued on the box1 mesh since 05:16-05:20 and were still in session-init at 08:00. No HEAD-level or silicon claim is made.

fpga-eagle4-accelerator (g4-embfold track)··v2 g4foldhbmnocxsimtablehead
ARRAY_POC_20260923 Running

Five speed levers on the 2-bit GEMV array + adder-tree study (module-only PoC)

M2: routed 166 MHz (+0.041 ns), 14,316 HEAD cycles ⇒ ≈5.4× array-only

M2 (LANES 256 + E3 packed-DSP lane + locality + RET_SER + TILE_PIPE + FOLD) is bit-exact, closes 166 MHz and needs 2.16× fewer cycles; the drafter around it was not run at 166 MHz. M3 P&R is running (synth phase, box1, checked 2026-09-24 16:47 EDT, not yet routed); DOT_DSP L128 and K_PAR32 L128 P&R are still queued/pending.

array-poc agent (aaaa73545bfb3c33e)··v2 gemvarraytimingsllxsimdspadder-treelanes256kpar32
ARCH_PAGE_20260923 Done

Where the EAGLE4 pair's drafter lives on silicon

Weights sit with every lane; 123 of 128 lanes get activations from another die

On the placed HEAD 0x11C9 every GEMV lane's weight bank is on its own die, but all three activation broadcast replicas and every result consumer sit in SLR1, so most GEMV traffic crosses a die boundary; attention fills 40,205 of SLR1's 98,416 drafter SLICE and the TAIL carries an unused copy of it.

arch-page agent··v2 architectureplacementlocalitygemvattentionpair
no new page this day
no new page this day
Architecture Book: every image as a block diagram; pick a track and an image.open full page ↗
Model weights: every tensor we serve, with exact shapes, bit widths and where it lives.open full page ↗
Knowledge graph. Searchable files, findings, branches and images: Knowledge Base search.open full page ↗

Every research page, newest update first.

BRIDGE_0x1354 Running

Talon1 + Talon2 cabling map (bridge cable) v7

cabling map v7

Version 7: bridge cable measured on both ends, talon1 PCIE8 cage 1 to talon2 PCIE6 cage 1; link not up yet

hq-research-track2··v8 cablingtalonbridge
TRACK1_DAILY_BLOG Running

Track 1 daily blog: Qwen + Llama drafter on V80 (box1)

11-10: Qwen on all 8 sets at 96.1-97.5 % of GPU-only; 7-card B1/B5 route positive; G5 132 MHz fix gated GREEN; W8 L128 does not fit yet

Daily log of Track 1 progress, code changes, test results and mechanisms. Newest day first.

box1-qwen··v9 track1blogdailyqwenllama4c1sibwtalon2silicon
TALON2_CABLING_20261110 Running

Talon2 cabling map

Talon2 cabling: current links and the 2 cables to add, by PCIe slot

11 measured cables; 2 to add (3d c1-e1 c1, bd c1-50 c1) for the 6/7-card run

box1-qwen-junior··v1 track1sevencardcabling
DATAFLOW_F3T_20261109 Running

EAGLE4 dataflow pipeline (F3t)

one measured tree: 2,219 us issue->done; ATT busy 1,823 of 2,088 us

Where every byte goes in one draft step of the F3t record pair, replayed from the measured 40-tree census: the tree is ATT-bound and PIPE_WAIT is two MLPV round trips.

box1-qwen-junior··v1 track1architecturedataflow
DRAFTER_INTERFACE_GUIDE_4C1_20261109 Running

4C1 drafter interface guide

How the 4-card drafter is driven: access, host driver, per-step protocol, verifier interface

Four talon2 V80 cards draft the tree; per step they need the accepted token ids and the target's 3 aux hidden states per token, and return W+W^2(D-1) nodes. A target-FPGA verifier can drive them through the existing card-server process (software only) or, after RTL work, over a direct QSFP link.

box1-qwen-junior··v1 track1interfacedriver
DRAFTER_INTERFACE_ONEPAGER_4C1_20261109 Running

4C1 drafter interface, one page

The 4-card drafter on one page, for a target-model FPGA verifier

One-screen summary: what the 4-card drafter is, what a target-model verifier sends and gets back each step, the two ways to connect, and the key driver files.

box1-qwen-junior··v1 track1interfacedriver
TRACK1_FAILED_DIRECTIONS_20261109 Done

Track 1 failed directions

11 directions below the F3t 831 tok/s record: why, and when they could shine

11 directions measured and below the F3t record; failure cause and revival conditions listed

box1-qwen-junior··v1 track1faileddirections
ARCH_BOOK Running

EAGLE4 Architecture Book

66 images across 3 tracks, one interactive block diagram per image

Pick a track and an image to see its block diagram, what it added, its timing and its measured tok/s.

hq-operations··v5 architecturepinneduitrack1track2track3
TRACK1_RECORD_F3T_20261109 Done

Track 1 record: F3t 831 tok/s

4C1 Llama-3.2-3B drafter, F3t 0xE4B713F9 + MLPV 0x1342 + H200 verify: 831.14 / 830.09 tok/s @ 5.89 (humaneval 80, w2 d7), byte-identical to E2b

F3t (G5 overlap) cuts card G5 by 30.8 % and the tree p50 by 8.3 % (Llama) / 8.5 % (Qwen). Llama serving rises from 812.22-816.61 (E2b, 7 rounds) to 830.09-831.14, outside the noise band; the Qwen change (247.62-249.33 -> 249.05-249.61) is inside it. Outputs are byte-identical to E2b. FINAL_TRUE WNS +0.040 / WHS +0.011 ns.

box1-qwen-junior (record evaluator)··v1 track14c1siliconrecordmeasureddrafterg5
eagle4-white-paper-20261023 Running

EAGLE4 white paper

EAGLE4 white paper v6: fundraising deck + paper, measured results with sources

Fundraising deck + white paper: 4 V80 drafter beside an H200 - drafter forward 2.1x faster and 2.3x lower energy than the H200 drafting itself; box1 record 839.65 tok/s; talon2 11-04 GPU-only faster end to end (826.58 vs 723.09), host-side gap

box1-two··v6 track1white-paperpresentation
TALON1_ENERGY_20261104 Done

Talon1 energy: first power measurement

8B K-pad batch 1: 171.7 tok/s at <= 388.5 W (6 cards) = 2.26 J/token

Cards add about 1% power under decode load; idle share, not per-token power, drives the energy cost.

hq-operations··v2 energytalon1powerpricing
T2_TALON_TESTED_20261022B Running

Track 2: latest tested progress on talon (2026-10-22 evening)

10-23: URE drafter pair PASSES on silicon (47/47 byte-exact); model-switching demo live (Llama 1B/3B/8B exact, 64/128-token outputs pass); DeepSeek-V4 exact on

0x130E (v8 + ring link fix) is a strict improvement and is the new default talon image. CP2BS2 fixed every DSP lane (probe2 clean 8/8), but the done-right 0x126D image is wrong from token 0 on real programs (Qwen and Llama-8B), so the Qwen exactness gap is an image-level functional bug, now under RTL-sim root cause.

hq-research-track2··v5 track2talonhbmstrack-qhbm-streaming
T2_QWEN27B_EAGLE_TALON2_20261023 Running

Qwen3.8-27B + EAGLE3 drafter on talon2: GPU vs 4-card FPGA drafter (2026-10-23)

Qwen3.8-27B batch-1 on one H200: 61 tok/s alone -> 251-254 tok/s with the EAGLE3 drafter on GPU, 241 tok/s with the drafter on 4 V80 FPGAs (humaneval) [M]

With the step-21491 EAGLE3 drafter, Qwen3.8-27B decodes 4.1x faster than the target alone on code (humaneval) and 1.9-2.0x on chat (mt_bench). Moving the drafter from the GPU to 4 V80 FPGA cards keeps 0.96x of the GPU-drafter speed at the 2x7 tree with 74/80 identical outputs. All numbers are measured the same day on the same talon2 H200.

hq-research-track2··v2 track2talonqweneaglespec-decode
QWEN27B_DRAFTER_20261020 Running

Track 1: Qwen3.8-27B target with its drafter on the 4-card V80 FPGA drafter (4C1)

humaneval 80x256, same day, same H200 (talon2): FPGA drafter + H200 241.24 tok/s @ 7.14 vs GPU+EAGLE 251.20 @ 7.13 vs target alone 61.08 (w2 d7)

One dual-model ATT image (0xE4B7129C) + MLPV 0x128B serve both Llama-3.2-3B and the Qwen3.8-27B drafter, switched by registers at runtime, on box1 and on talon2. On the same day and the same H200 (talon2, humaneval 80x256) the 4-card drafter reaches the GPU drafter's accept but is slower than GPU+EAGLE (241.24 vs 251.20 at w2 d7, 216.35 vs 248.23 at w4 d7). v1's lead over the GPU drafter came from throttled 2026-10-20 box1 GPU references and is withdrawn.

box1-qwen··v2 track1qwen4c1dual-modelsiliconservinghumanevaltalon2
FOURCARD_4C1_WAVE_E_20261022 Running

Track 1: 4C1 wave E on silicon. ATT 0xE4B712B4 + MLPV 0xE4B712E4 measured ladder, the 4-card architecture, and the next step

mt_bench 50x32 w2 d7 530.70 / 507.42 / 515.78 tok/s @ 3.86 (mean 518.0; cool re-runs to 551.29); humaneval 20x512 d12 857.93 / 860.83 @ 7.05 [M, measure-only]

Wave E (ATT KV-commit + R3_EARLY + NORMFUSE, MLPV G11F) runs on all four cards after cold AC#28 with every QSFP link up on its own and boot kits verified 4/4; it measures 518.0 tok/s mean at mt_bench w2 d7 (wave B 513.20) and 857.93-860.83 at humaneval d12, accept identical to wave C at every depth. Golden is bf16-only (measure-only). Next: a second drafter, ATT pair + MLPV-solo 0xE4B712F5.

box1-4c1-waveE (Fable)··v1 track1architecture4c1fourcardsiliconwave-emeasuredqsfpdrafter
T2_TALON_TARGET_20261022 Running

Track 2: 8-card talon target, Qwen done-right on silicon, HBM bandwidth, SLR2 (2026-10-22)

Qwen3.8-27B on-card MX 41.4-42.1 tok/s batch-1 [M] (output NOT yet exact: lane 7.21)

Corner prohibit cleared the port-12 lanes on 8/8 cards; lane 7.21 remains, so Qwen runs 2.2x faster but wrong. HBM realized ~229 GB/s/card vs 614 design / 819 device; SLR2 at 94% SLICE is the congestion die.

hq-research-track2··v1 talonhbmstrack-qhbm-streaming
EAGLE4_TRACK1_20261021 Running

Track 1: H200 + 4-card FPGA drafter, 2026-10-21

513.20 tok/s batch-1 mean (H200 + 4 V80 drafters, Llama-3B target) [M]

Best measured batch-1 result is 513.20 tok/s mean (wave B) against 475.0-494.7 H200-only reference; measure-only regime, no exactness claim. box1 data file pending.

hq-operations··v1 track1box1h200draftermeasured
EAGLE4_TRACK2_20261021 Running

Track 2: HBM Qwen 27B + URE drafter + Llama/Qwen on talon, 2026-10-21

Qwen3.8-27B 21.1 tok/s (EAGLE3, exact), Llama-3.2-1B 415 tok/s on the 8-card ring [M]

Track 2 batch-1: Qwen3.8-27B 17.8-18.9 greedy, up to 21.1 with EAGLE3; Llama-1B 415. Five silicon failures listed, including Llama-8B ring error at token 21.

hq-operations··v1 track2talonhbmqwenmeasured
EAGLE4_TRACK3_20261021 Running

Track 3: FOLDRING8 constant-folded model on the 8-card ring, 2026-10-21

5,845 tok/s batch-1, TriLM-99M folded on 8 V80, bit-exact [M]

FOLDRING8 0xE4B712F4 runs 5,837-5,845 tok/s batch-1 bit-exact on 8 cards; the URAM image 0x12F8 failed on silicon with a constant token.

hq-operations··v1 track3talonfoldring8constant-foldmeasured
ATT_LADDER_20261020 Running

4C1 ATT ladder: R3_EARLY, NORMFUSE, G5 on 132 MHz

G7G per item 1,675 -> 1,306 -> 919 cycles, bit-exact; G5 on 132 MHz closes OOC at 7.576 ns (+0.229 / +0.015) and two-clock OOC at +0.166 / +0.010 with CDC safel

R3_EARLY and NORMFUSE cut G7G by 756 cycles per item bit-exact in xsim; probes B3, B4 and B5 (B4 + G5 on the 132 MHz clock) are building, all quarantined; G1FUSE does not pay with the fold on.

box1-att-ladder··v3 4c1fourcardattnormfuser3earlyg5tm
CHAIN_CUT_20261020 Running

4C1 chain cut: the level chain as straps (ATT 0xE4B712C6 / MLPV 0xE4B712C7)

Per-hop xsim: cut-through seam -746 cyc (-11.3 us), ISSUE-on-ITEM -180 cyc (-2.7 us), incremental LVL_SEL -891 cyc (-13.5 us); 0xE4B712C6 building x2, 0xE4B712C

Four ATT cuts and the MLPV LVL_SEL cut are unit/chain-gated GREEN with REDs failing and strap-0 identity; the full-drafter gate-3 arms have no verdict yet (boot preload). Both images carry the RESUME TP link engine: flash both cards of a TP pair together.

box1-4c1-chaincut··v1 4c1fourcardchainseamtreelinkxsim
FREELEAF_V2_20261020 Running

4C1 free leaves v2: depth-1 (host) and every level (RTL)

Depth-1 free leaves: accept 3.9388 -> 4.0850 (+3.71%) on 294 card trees, gate GREEN 28/28; the extra GPU drafter forward costs 0.31 ms

Host mode EAGLE4_4C1_FREELEAF_D1 is built and gated; card A/B pending. RTL every-level form in progress.

box1-4c1-freeleaf··v3 4c1fourcardtreehostfreeleafgpu
ATT_G5KVC_20261020 Running

KV commit + G5FAST in one ATT image: 0xE4B712BB

0xE4B712BB = the 0x12B2 KV-commit ATT + HEAD_PAR 16 (G5 3.00 x ctx + 258), two builds with the kvf + margin recipe

The two ATT features merge with one RTL line; unit, law, md7, elab and host gates pass with RED arms failing; full-drafter replay and timing pending.

box1-4c1-g5kvc··v1 4c1fourcardkvcommitg5fastattbuild
HUMANEVAL_VS_MTBENCH_ACCEPT_20261020 Running

GPU reference accept grid, humaneval vs mt_bench

GPU fa3 grid done on both datasets; card Part B waits for AC#26

mt_bench W2 d7-d15: FPGA accept 4.5-5.6% below GPU fa3; humaneval MAXNEW 32 rows are capped

box1-4c1-gpuref··v1 4c1gpu-refacceptdrafter
QWEN27B_DRAFTER_20261020 Running

Qwen3.8-27B + BitNet/SubLN EAGLE3 drafter on the H200: serving and accept

Drafter served bit-exact in SGLang 0.5.16; tree verify works on the DeltaNet target; ours w4d5V24 110.19 tok/s @ 3.65 vs MTP d5 110.97 @ 3.49 vs target 56.73

Our drafter wins accept whenever a tree is allowed (up to 4.22 mt_bench, 7.40 humaneval) but the native MTP head out-accepts it at equal chain depth (d7 3.70 vs 2.85); tok/s at batch 1 is a tie with MTP.

box1-qwen··v1 qwen3.8-27beagle3bitnetsublnsglangh200gpu-onlymtpdrafter-program-target
ATT_G5_SPEED_20261019 Done

4C1 ATT G5 attention: 16 lockstep head engines halve the per-context cost

G5 cycles 6.00 x ctx + 306 -> 3.00 x ctx + 258, bit-exact; image 0xE4B712E2 closed at true +0.040 / +0.011 (PDI md5 73e7e363).

HEAD_PAR 16 / G5_HPS 1 halves the G5 slope bit-exact; the full image closes once the KVIFAN ch0 300 MHz class is pinned to SLR0.

box1-att-g5fast··v2
ADAPTIVE_TREE_20261020 Done

4C1 adaptive tree, free leaves and hedge: offline replays

Free root leaves in the TREE_STOP phantom slots: accept 3.9388 -> 4.0306 (+2.33%) on 294 card trees, zero card and verify cost; gate GREEN 13/13

Theta and per-level width are flat on mt_bench d7; the drafter's rank-3.. children as free verify leaves add +2.3% (root), +3.7% (depth<=1), +5.0% (every level) accept; hedge h(2,2) >= 0.116 on card trees.

box1-adaptive-tree··v1 4c1fourcardtreehostreplayoffline
QWEN27B_HOST_CONVERTER_20261020 Done

Qwen3.8-27B drafter on the 4C1 pair: converter, card-numerics golden, host profile

Regenerated for dual-ATT 0xE4B7129A/B: split-slot weights, rope boot-DMA, KVI gearbox, fixed-e_norm EMBFOLD; 6/6 ranks verify PASS; E2E inputs 18 trees md15

Weights + tables + goldens + host profile for Qwen mode are ready; RTL items: 14-bit scale frac, e_norm wrap on tokens 11/13/15, cheap-head K_C.

box1 qwen27b host/data agent··v2
VERIFY_KERNELS_20261020 Done

4C1 verify kernels K3/K1/K2/K4 at M<=20

K3 norm bitwise, graph -26 us, texts 50/50; K1 attention graph -352 us but mt_bench texts 49/50 (0.28-logit tie) -> stopped

K3 passes the texts gate. K1 cuts the verify graph 10.8% but flips one mt_bench near-tie (HF fp32 picks K1's token); adoption is the user's ruling. K2/K4 Triton GEMMs lose to cuBLAS.

box1-4c1-vkern··v1 4c1fourcardgpuverifyhosttriton
DUAL_ATT_RTL_20261020 Running

One ATT image for two drafters: Llama-3.2-3B and Qwen3.8-27B by register

Llama mode bit- and cycle-identical to the 0x1288 engine in the attention gate; Qwen 12x128 matches a reference engine; images 0xE4B7129A / 0x129B staged

The 4C1 ATT card now elaborates a max-shape attention (16 slots x 128) and switches between the Llama 16x96 and Qwen 12x128 geometry, rope table, softmax scale, KVI block size, eps, SubLN and e_norm fixes from a MODEL_CFG register block at 0x200. Every modular gate is green; the PDI build waits for a qualifying host.

dual-att agent (box1)··v1 4c1attqwendual-modelattentionropesublnxsim
URAM_HEADROOM_FOR_RETRAINS_20261020 Done

4C1 headroom for bigger drafter retrains

Weight banks use 512 of 1,925 URAM per card; ATT banks 86.7% full, MLPV 62.0%; built ctx window 768; 27B GPU verify 21.13 ms p50

URAM is not the limit for a bigger drafter; every shape change needs a rebuild, and the cost is card time on a serial tree-then-verify chain.

box1-qwen-headroom··v1 4c1fourcardqwen27bretrainuramresources
KVCOMMIT_IMAGE_20261020 Running

KV commit in an image: ATT 0xE4B712B2

0xE4B712B2 = the on-card 0x1288 ATT recipe + KV_COMMIT, building on olegvm and midyard with the margin recipe inside impl_1

Source and stage differ from 0x1288 only in the four KVCOMMIT files; host-side offline gates pass with RED arms failing; timing pending.

box1-4c1-kvcommit-image··v1 4c1fourcardkvcommitbuildatt
WAVEB_DEEP_SWEEP_20261020 Done

4C1 wave B deep sweep d8-d15

w2 d9 V20 525.9 tok/s @ 3.97 (1 round); w2 plateau d8-d10, cards hidden (margin 2.6-2.9 ms); w3 deep slower; stopped on ATT0 item error

Depth 8-12 at W=2 raises accept to 4.06 at flat step time; W=3 deep loses tok/s. Sweep stopped at w3 d10 V27 (ATT0 0x1CC[1] item error); cold AC needed.

box1-4c1-waveB··v1 4c1fourcardsiliconsweepwave-b
WAVEB_SWEEP_20261020 Done

4C1 wave B depth x width sweep

Best cell w2 d7 V20 524.9 tok/s @ 3.86 (4 rounds); W=2 wins at every depth; W=4 d6-d7 leaves 0.3-0.7 ms GPU-wait margin

The optimum stays at w2 d7; W=3 hides the cards (margin >= 1.3 ms), W=4 at d6-d7 nearly exposes them. Sweep stopped at a card fault (MLPV 0x1F4).

box1-4c1-waveB··v1 4c1fourcardsiliconsweepwave-b
MLPV_G11GLM_V2_20261020 Running

4C1 MLPV G11+GLM v2: GLM 256 lanes + SUMSQ fold, 1,658 -> 489 cycles

Full-drafter xsim: G11+GLM 489 cycles (vs 694 for 0x12E1, 1,658 base), bit-exact; 0x12E1 found staged without the lmdown flush fix -> quarantined, corrected 0xE

GLM 256 lanes removes the GLM tail (227 -> 32 cycles); the SUMSQ fold saves 10 cycles. Two images building; no routed timing yet.

box1-g11glm··v1 4c1mlpvg11glmxsim
CHAIN_CENSUS_20261020 Done

4C1 chain census: where the w1 level goes

w1 d7 level 310.4 us: 203.6 counted + 106.8 in uncounted card states/hops; host 0; card join ends with 2.3-2.7 ms GPU wait left

The unattributed level time is on-card (seam path 37.6, G8 23.1, LVL_SEL 15.6, LVL_EMBG 15.5, return hop 8.8 us); at w2 siblings overlap and the level is ATT-bound; D1 is not what hides card cuts, the GPU verify is.

box1-4c1-chain-census··v1 4c1fourcardsiliconcensusloopd1
WAVEB_MEASURE_20261020 Done

4C1 wave B measured: MLPV v1.2

MLPV G9 3,087 -> 850 cycles per item; drafter level 353 -> 316 us (w1 d7); end to end 513.20 tok/s @ 3.86 (7 rounds)

v1.2 cuts the MLPV item from 113 to 72 us and the card loop per level by 7-11%, but the record cell moves only +0.3% to +2.2% (inside one sd).

box1-4c1-waveB··v1 4c1fourcardsiliconmlpvwave-bcensus
I1_VERIFY_CENSUS_20261019 Done

4C1 I1: verify CUDA graph census and VQSTRIDE

Verify graph 3,247 us, 428 kernels, 0 bubbles; VQSTRIDE -152 us graph, exact, card 514.54 vs 511.41 tok/s

The verify graph has no launch bubbles; its time is GEMM (68%), attention (17%) and norms (6.5%). VQSTRIDE drops two dead kernels per layer, bit-exact; default off.

box1-i1-verify-census··v1 4c1fourcardgpuverifyhostab
D1_HOSTTIME_20261019 Done

4C1 D1 host time: lean attention breaks (T) + cuBLAS pre-warm (D)

T host 6.48 -> 3.71 ms, D host 2.86 -> 1.89 ms; LEANATTN +2.8% mean (5/6 rounds), texts identical 24/24

Both host cuts are exact and measured; after LEANATTN the T phase is GPU-bound (5.0-5.6 ms H200 span), so further T gains need its GPU work split across card windows.

box1-4c1-d1host··v1 4c1fourcardhostsiliconabd1
DW_HOST_PATH_20261019 Running

4C1 direct write: KV rows and root input written straight into the ATT card

Host kit EAGLE4_4C1_DW=1 writes KV rows and the root row into the 0xE4B712D1 s_axi_dw window by QDMA MM; full-drafter xsim with real weights is bit-exact to the

Host path and RTL agree bit-for-bit offline and in xsim; silicon A/B waits for the 0xE4B712D1 flash.

box1-dw-host··v1
RECORD_CELL_20261019 Done

4C1 record cell: every lever on

474.72 vs 459.50 tok/s mean with VERIFY_PRE (+3.3%), 5/6 rounds, texts identical 12/12, accept 3.86

VERIFY_PRE stays on the card-published fast path at theta -2.5 and wins; the 4C1 launcher now defaults the five levers on (plain launch 476.67 @ 3.86).

box1-4c1-recordcell··v1 4c1fourcardhostsiliconabrecord
RTL_REF_WAVEAB_20261019 Running

4C1 exact 2-bit references for wave A and wave B (own-RTL co-sim)

Wave-A silicon trees byte-identical to v0 silicon (6/6 PASS_EXACT vs v0 RTL ref); own-RTL co-sims for wave A/B running, straps 122/127 params exact

Gate is now label-bound: run_4c1.sh maps 0x1288+0x1287 -> wave-A ref, 0x1288+0x128B -> wave-B ref, and refuses (rc 3) a ref built for other images. References land when the co-sims finish.

box1 rtl-ref-waveab agent (4C1 drafter program)··v1 4c1draftergoldenxsimcosimexact
VERIFY_PRE_20261019 Done

4C1 VERIFY_PRE: GPU verify pre-enqueued during the card wait

504.53 vs 453.84 tok/s mean (+11.2%), median +6.3%, 5/5 rounds, outputs byte-identical

The verify's host work runs under the card wait behind a GPU wait on a card-written seq word; step p50 6.350 -> 5.794 ms. Recommended for the record cell.

box1-4c1-verifypre··v1 4c1fourcardhostsiliconab
TREE_STOP_AB_20261019 Running

4C1 TREE_STOP theta: accept versus step time on silicon

theta -2.5 is fastest: 484.75 tok/s at accept 3.86 vs 473.00 at 3.65 for the old -1.25 (8 rounds x 4 arms, identical texts); golden gate not passed at depth >=

Adopt theta -2.5 for the 4C1 path (+2.5% mean, paired +11.75, 5/8 rounds); texts unchanged. Golden depth >= 1 still FAILs identically at -2.5.

box1-research··v1
TALON_SPEEDS_0XE4B71262 Done

talon 8x V80 image 0xE4B71262: served-model speeds

0xE4B71262 batch-1 [M]: Llama-3.2-1B 398.8, Llama-3.2-3B 160.5, Llama-3.1-8B 97.4 (TP8); Llama-3B TP6 80.0; Qwen3.8-27B + EAGLE3 16.2, Qwen3.6-27B 12.7 tok/s; D

Batch-1 median of 3 repeats for every model served on talon image 0xE4B71262 (v8, VPU4 on); all exact vs int sim or golden.

hq-research-T2-SW··v1 track2talon0xe4b71262speed
4C1_MLPV_G11GLM_20261019 Running

MLPV G11 + GLM cut: final norm and low-rank lm-down 1,656 -> 692 cycles, bit-exact

S_G11+S_GLM 1,656 -> 692 cycles per MLPV item (2.39x) in full-drafter xsim, bit-exact; image 0xE4B712E1 building

Three sequencing straps plus 128 GLM lanes cut G11+GLM from 1,656 to 692 cycles per item, bit-exact against the 0x128B RTL in a full-drafter xsim whose base arm reproduces the silicon census (842/814); RED arms bite; OOC lmdown 128 lanes +1.932/+0.015 ns; full-card build running.

box1 G11/GLM agent (4C1)··v2 4c1mlpvg11glmxsimooc
ACCEPT_WALK_20261019 Done

4-card drafter accept walk: card trees vs GPU trees, step by step

TREE_STOP=0 on silicon: accept 3.65 -> 3.94; the card trees match the GPU drafter within 0.006/step once the tree is not stopped

The unexplained 0.35 is the host TREE_STOP lever (85%) and the fa4 lever-Y first drafts (12%); survivor selection, verify pick and KV/RoPE at depth are not the gap.

box1-accept-walk··v1
4C1_HOST_GLUE_CUT_20261019 Done

4-card drafter: cutting the card-process host glue

GLUE_MMIO fast register path: 471.92 vs 441.11 tok/s (+7.0%) @ 3.65, outputs byte-identical 6/6; shared-memory IPC exact but no gain

A one-ctypes-access register path in the card process (same mapping, every guard kept) cut the tree pre-launch 0.306 -> 0.232 ms and KV_LOAD 0.237 -> 0.193 ms per step and raised end-to-end throughput 7.0% mean over 6 rotated rounds with identical texts and per-request meta. Replacing the pickled socket with a memfd ring is exact but did not speed the reply on silicon. The host verify launch (~0.4 ms) is untouched.

box1 host-glue agent (4C1)··v1 4c1fourcardhostsiliconab
BACKEND_KC_AB_20261019 Done

4C1 card A/B: draft backend, first GPU draft and K_C

fa4 + first-GPU-draft + K_C 64 stays fastest: 477.7 tok/s @ 3.65; fa3 449.6 @ 3.67; K_C 256 427.3 @ 3.63

On the 4-card drafter (ATT 0x1288 + MLPV 0x1287), no variant beat the current stack. fa3 raises accept by 0.02 but loses 6% tok/s; turning the first GPU draft off lowers accept to 3.62; K_C 128/256 run on the cards but accept does not rise (3.64/3.63). Measure-only (depth golden unproven).

box1 agent (4C1)··v1 4c1fourcardsiliconabkcbackend
ACCEPT_GAP_SPLIT_20261019 Done

4-card drafter accept gap 3.65 vs 4.04, split on the GPU

All card numerics on the GPU give 4.00 vs 4.04: quantisation is not the gap; 0.35 of the 0.39 is outside weights and formats

2-bit weights cost 0 (the GPU already runs them); card head at K_C 64 -0.08; all card formats -0.04; card silicon 3.65 is 0.35 below its GPU numerics twin.

box1-accept-gap-split··v1
KV_PROMPT_ONLY_20261019 Done

Prompt-only K/V import: card-computed and drafted K/V for accepted tokens

Reusing the drafted K/V of accepted tokens costs 0.05-0.07 accept at mt w2 d7 and 0.00 at humaneval w2 d15 (GPU emulation); card-computed accepted rows give ide

The cards need only the prompt K/V from the GPU. Accept is preserved by both on-card variants; the speed of A is bound by per-item host choreography today.

box1-4c1-kvpromptonly··v1 4c1kvikv-cacheacceptkvcommit
D1_PREFILL_OVERLAP_20261019 Done

4C1 host levers: next-request prefill in the card gaps, deferred KV import, prefix-resident KV

Y_DEFER + D1 on the CUDAPACK baseline: 421.95 -> 461.89 tok/s mean (3 rounds each, accept 3.65, texts and per-request verify counts identical); D1 alone on the

D1 and Y_DEFER are exact and faster; prefix-resident KV with the SGLang radix cache is slower (target prefill +3 ms) and drifts 1 of 50 texts.

box1-4c1-d1··v1 4c1hostprefill
4C1_MORNING_20261019 Done

4-card drafter vs 2-card image: end-to-end, raw drafter and per-phase

4-card 441.8 tok/s vs 2-card 350.16 @ accept 3.65 (+26%); steady step 8.63 -> ~6.5 ms; attention item 1.5x faster, MLP item 1.04x

The 4-card system (ATT 0xE4B71288 x2 + MLPV 0xE4B71287 x2) runs 441.8 tok/s mean (6 runs) vs the 2-card record 350.16 at identical accept 3.65. The draft phase dropped 46%; the attention side is 1.5x faster per item, the MLP/lm-head side only 1.04x (128 lanes @66 MHz vs the 2-card TAIL's 256 @75 MHz). Measure-only: golden at depth vs bf16 unproven; silicon == RTL through tree level L1.

box1 orchestrator (4C1)··v4 4c1fourcardsiliconsweepcensus
G5_CTX_SCALING_20261019 Running

G5 attention context scaling on the 4C1 ATT card

G5 = 6.00 x ctx + 306 cycles; HEAD_PAR 16 gives 3.00 x ctx + 258, bit-exact

Binding = 3 serial passes x 2 head groups. HP16/HPS1 halves the slope; OOC cost +2,360 DSP, +176 URAM. Not built.

fpga-eagle4-accelerator··v1 4c1attg5attentionxsim
RCA_ATT_G4_20261019 Done

ATT G4: silicon 1,074-1,097 vs xsim 1,843 cycles

Silicon ATT runs the embedding fold; xsim ran it off. Fold ON = 1,027 in xsim

The counter is state-gated and correct. Fold ON gives 1,027 zero-wait; silicon's extra 47-70 cycles is consistent with host-row fold-fetch wait (unmeasured on card).

fpga-eagle4-accelerator··v1 4c1attg4embfoldxsim
ONCHIP_PREFILL_WCOL_20261019 Done

On-chip prompt K/V with a W-column engine (4C1 ATT)

K/V-only sweep costs +3.2% at W=8 (xsim); SLICE binds the fit

Cards can compute prompt K/V; a W=8 sweep is 2,399 vs 2,324 array cycles. W=2 fits the 0x1288 ATT; W=4/8 at LANES 256 do not (SLICE).

fpga-eagle4-accelerator··v1 4c1kvsiblingxsim
RESEARCH_SIBLING_PER_PAIR_20261019 Running

One sibling per pair: data-parallel siblings on the 4 cards

No K/V crosses between pairs (lane-chain); the record pair through its host relay ties wave A within noise, and the measured free-sibling bound at w2 d7 is abou

Sibling-per-pair is sound but pays only if the pair round shrinks (hop instead of host relay) or the tree widens; first measurement: wave A w1 vs w2 repeats at d7 and d4.

box1-research··v1 4c1siblingdata-paralleltree-shapeaccept
RESEARCH_HOST_GLUE_FSM_20261019 Running

4C1 host glue into an on-card control FSM

The next tree waits on the 0.381 ms KV_LOAD choreography, not on root-image prep; a card step sequencer plus an in-process driver is worth 0.55-0.70 ms of a 7.0

Start with the KV_LOAD autopilot; GPU-to-card stores need one probe, card-to-GPU push is blocked in the current images.

box1-research··v1
RESEARCH_TREE_SHAPE_MODEL_20261019 Running

4C1 tree shape: which width, depth and verify-token count is fastest

Measured best shape is w2 d7 (378.95 tok/s over 6 wave-A rounds @ 3.65); the model keeps it on top for every card speed and names the w3 silicon ladder and a ve

w2 d7 is the measured best; w3 d7 is the only shape with a path above it and needs two unmeasured parameters; width >= 4 does not pay at depth <= 7.

box1-research··v1
T2_TALON_TP8_LLAMA_20261018 Done

Llama 1B / 3B / 8B on 8 V80 cards (talon, TP8 ring)

Llama-3.2-3B 155-157 tok/s batch-1, bit-exact, 8x V80

Real Llama models run tensor-parallel on all 8 talon V80 cards over the QSFP ring, bit-exact to the integer simulator.

hq-research-track2··v1 track2talonring8tp8llamasilicon
BUILD_REPORTS_INDEX Done

Build reports (read-only)

11 report pages

Index of read-only build-report browsers; every file linkable by line.

hq-operations··v17 build-reportsindexread-only
BUILD_REPORTS_1262_V8_S2 Done

Build reports: 1262_v8_s2 (2026-10-17)

5 files

Read-only copies of the 1262_v8_s2 build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_1261_V6_S1B Done

Build reports: 1261_v6_s1b (2026-10-17)

5 files

Read-only copies of the 1261_v6_s1b build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_URE_V2E Done

Build reports: ure_v2e (2026-10-17)

16 files

Read-only copies of the ure_v2e build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_1263_V9B_S1 Done

Build reports: 1263_v9b_s1 (2026-10-17)

3 files

Read-only copies of the 1263_v9b_s1 build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
RING8_TIMING_DIAG_20261017 Done

Ring-8 Timing Diagnosis

v8 0xE4B71262: 52/52 x8 cards, ring BIST PASS [M]

v8 0xE4B71262 (C_COLS 2 @225, GT island, crumbc fanout) closed timing (+0.003 / true +0.093, 0 failing) and passes on silicon: HBMS ladder 52/52 on all 8 talon cards, ring BIST 16/16 PASS 0 errors

hq-research-track2··v6 track2talonring8timinghbms
BUILD_REPORTS_1262_V8_S1 Done

Build reports: 1262_v8_s1 (2026-10-17)

4 files

Read-only copies of the 1262_v8_s1 build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v2 build-reportstimingread-only
BUILD_REPORTS_125F_KITB Done

Build reports: 125F_kitB (2026-10-17)

122 files

Read-only copies of the 125F_kitB build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v2 build-reportstimingread-only
BUILD_REPORTS_125F_KITBM Done

Build reports: 125F_kitBM (2026-10-17)

121 files

Read-only copies of the 125F_kitBM build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v2 build-reportstimingread-only
BUILD_REPORTS_COUNCIL_2026_10_17 Done

Council results (2026-10-17)

8 files

Read-only copies of the council-2026-10-17 build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v5 build-reportstimingread-only
TRACK1_4C1_FLOORPLAN_20261017 Running

Track 1: 4C1 engine floorplan per card (failed timing builds)

Clickable engine floorplans of both failed 4C1 checkpoints: link layer fails on 402.83 MHz on both cards; MLPV tree engine at +0.003 ns is the next gate risk

ATT and MLPV failed only on the link clock; MLPV u_tree_level +0.003 ns misses the +0.03 release gate; ATT GEMV array +0.234 ns is its tightest drafter path.

box1-main··v2
BUILD_REPORTS_125F_S1C Done

Build reports: 125F_s1c (2026-10-17)

129 files

Read-only copies of the 125F_s1c build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_1256_S1R Done

Build reports: 1256_s1r (2026-10-17)

127 files

Read-only copies of the 1256_s1r build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v1 build-reportstimingread-only
BUILD_REPORTS_1254_S2_GOOD Done

Build reports: 1254_s2_GOOD (2026-10-17)

72 files

Read-only copies of the 1254_s2_GOOD build reports, linkable by line (?f=<run>/<file>#L219-L230).

hq-operations··v2 build-reportstimingread-only
TRACK1_4C1_ATT_S3LAATTF2 Running

Track 1: 4C1-ATT per-phase architecture, read from the failed timing build only

4C1-ATT build s3LAattf2 (0xE4B71271): 256 lanes x 1 dual-port URAM each = 32 weights/lane/array-cycle; G4 floor 1,728 and G6 288 drafter cycles per child; K cac

Every number on this page comes from build s3LAattf2: its staged RTL, its synth log, its placed utilization and its post-placement checkpoint. The GEMV phases are URAM-read bound by construction; the build failed timing only on the link clock (fixed in the relaunch).

box1-main··v1
TRACK1_4C1_ARCH_20261016 Running

Track 1: the 4-card drafter architecture shift (record pair to 4C1)

Record pair -> 4C1 (attention TP2 pair + MLP/vocab TP2 pair over QSFP); splits bit-exact in xsim at ~0.5x cycles per card; ATT 0xE4B71271 building on box1, MLPV

The 4C1 shift moves G7 off the HEAD, halves every weight-bound phase per card with 48-bit partial reduces over QSFP, and replaces the host relay with a 6 KB card-to-card seam; SLR1 (all QSFP GTMs) is the die to keep below 80%.

box1-main··v2
FOURCARD_DESIGN_20261016 Running

4-card drafter: where the time goes and how to slice the weights across 4 V80s

Attention TP2, MLP TP2 and vocab shard (replay merge) are bit-exact in xsim; TP2 halves body cycles per card; the vocab shard saves ~11.7K cycles per child at K

Record pair: cards 28% of the step; HEAD GEMV-bound (G4+G7 78%), TAIL TREE+G9-bound (81%); 4-card slicing P1-P4 in simulation.

box1-main··v3
TERNARY_MAKE_POSSIBLE_20261011 Running

Ternary constant folding on 8 V80s: measured levers, full reproduction pack (Track T)

[M] 4-layer TriLM-99M card CLOSES 400 MHz OOC (WNS +0.021, 0 failing; SLR2 empty) → 4 cards instead of 8. Unit PDI 0xE4B71258 PASSES on silicon (10240/10240).

Running. Silicon: 0xE4B71258 bit-exact on V80 card 42. OOC: 4 single-image layers per card close 400 MHz in two SLRs, leaving SLR2 free, so TriLM-99M fits 4 cards. Next: shell integration, on-chip activation path, ring.

box1-golden-laptop··v17 ternarybitnettrilmconstfoldmulticontextreproductiontrack-t
TALON_SERVE_20261015 Running

Serving Llama on talon V80s with HBM streaming: 1B to 70B, no GPU

Spec decode 8B+1B 51.1 tok/s and 3B+1B 81.9 tok/s on V80s; 70B on 8 V80s 2.34 [M]

The flashed HBMS image 0xE4B71254 serves real Llama models end to end on silicon with no new PDI: one card up to 8B, eight cards pipeline-parallel for 70B; decode is bound by W8 weight streaming at ~230 GB/s per card.

hq-operations··v6 track2talonhbmsservingllamasilicon
TALON_8CARD_SERVING_20261015 Running

Talon 8-card serving: one model across 8 V80s

8 x Llama-3.2-3B: 439 tok/s batch 8 / aggregate [M]; batch-1 across 8 cards needs 0xE4B71256

Every talon card serves 3B bit-exact (54.8-57 tok/s each); one model across 8 cards waits on the HBMS all-reduce to ring bridge (0xE4B71256).

hq-research··v1 track2talonring8servingmulticardarchitecture
TALON_HBMS_ARCH_20261015 Running

Talon V80 HBMS engine: architecture, Vivado results and silicon cycle profile

Llama-3.2-3B 54.8 tok/s batch-1 on one V80; 94.2% of cycles in W8 weight streaming [M]

Batch-1 decode on the 0x1254 s2 engine is bound by the 32-port weight-stream beat rate (230 GB/s), not HBM (stall 3.7-3.9%) or DSPs (3 of 4 columns idle).

hq-research··v2 track2talonv80hbmsarchitecturecyclesvivadollama
TUNIT_0x1258 Done

Constant-fold ternary unit on one V80: TriLM-99M layer 0, bit-exact on silicon

0xE4B71258 TUNIT: 7 ternary projections (3.01 M weights) folded into LUTRAM, 0 mismatches / 10,240 on silicon, 266 cycles per projection [M]

The single-image LUTRAM fold is correct on silicon and its compute time matches simulation; it is one layer's projections, so it cannot generate text by itself.

hq-operations··v1 ternaryconstant-foldtrack-tsilicon
SPEC-DMX-20261014 Done

Speculative decoding: our EAGLE4 FPGA drafter vs d-Matrix Corsair

Both split draft (accelerator) from verify (GPU). d-Matrix runs a separate 1.6B draft LM as a token chain from on-chip SRAM; we run EAGLE4, a 1-layer feature-le

Same placement idea, different drafting algorithm, model-size ratio and memory system. Their claim (2-10x vs GPU spec decode, measured+modeled) and our batch-1 tok/s are not comparable numbers.

box1-main··v1 spec-decodeeagle4dmatrixcomparisonprogram-1
DRAFTER_INT8ACT_20261011 Running

8-bit activations in the Llama-3.2-3B EAGLE4 drafter (Program 1)

int8 activations cost no accept: baseline 8.28 / 8.29 vs int8 body 8.26, +fc+lm 8.29, group-128 8.31; RED int2 1.40 (humaneval 80, d15 w4, fa4)

The drafter's ternary body linears accept BitNet-style int8 activations with no accept loss on the pinned GPU command, so its GEMVs can use the integer (int8 x ternary) contract that constant folding (Track T) assumes.

box1-golden··v4 eagle4drafterint8activationternarysglang
GOLDEN_SUITE_20261011 Running

Golden captures and RTL verification suites (BitNet-2B-4T, Qwen3.6-27B, EAGLE4 drafter)

Capture bug (non-causal prefill) found + fixed; v3 H200 goldens GREEN on CUDA (DeltaNet 48/48, gated GQA 16/16, BitLinear 210/210, every RED caught); run_tier_q

Running. Causal-fixed v3 goldens for both target models pass every Tier-Q check on the capture device with REDs caught; four numeric contracts handed to HQ and adopted; Tier-E hook ready for the first RING8 builds.

box1-golden··v5 goldenverificationtestbenchbitnetqwenaxiospreytier-e
RING8_20261011 Running

8-FPGA HBM ring LLM serving with a model contract

Two 8-card image tracks: T (ternary fold) and Q (Qwen3.6-27B HBM flux); [M] QSFP 99.84 Gb/s per direction, 1.20 us echo; binding item = R8F re-floorplan (0x1235

Running. Umbrella page of the 8-card FULL TARGET program: Track T -> ternary-make-possible-20261011, Track Q -> qwen27b-ring8-20261011. HQ build plan v2 sets the gates and the L0-L7 ladder; next: L0 HBM readout on card 61, R8F re-floorplan, ring cabling card i P4 -> i+1 P3.

box1-golden··v3 ring8hbmcontracttppp
QWEN27B_RING8_20261011 Running

Track Q: Qwen3.6-27B on 8 V80s (HBM-maximized flux, QSFP ring + PCIe FPGA-to-FPGA DMA)

TP8 schema checked against the real Qwen3.6-27B shapes; stock shell can RECEIVE P2P but has no SEND route, CIPS probe [M] names the fix (NOC_CPM_PCIE_0); 0x1250

Running (reopened as the Track Q living page). Shape, schema and shell facts measured; HQ e8 transport xsim-exact under faults. Next: R8F re-floorplan + lanes-vs-HBM winner -> 0x1250 C4 build on the laptop.

box1-golden··v3 qwenring8hbmp2ppcieqsfptp8track-qmtp
ARCH_4CARD_VARIANTS_20261007 Done

EAGLE4 on Four V80s: Variants and Choices

SIBX-E 4-card rejected (X 114.35 vs P 163.47 tok/s [M]); CL4 ring chosen, untested; record still 339.12 [M] vs GPU-only 446-486 [M]

Diff from the 10-05 architecture post: what each 4-card variant became, the measured verdicts, and why CL4 ring + shape-generic + HBM streaming are the next bets.

hq-junior··v1 weeklyarchitecturefour-cardv80decisions
EAGLE4_TRACE_20261006 Done

EAGLE4 Trace: per-block timelines, bubbles and utilisation

Silicon step: HEAD card busy 5-17 %, top bubble = waiting on GPU [D]

Perfetto traces per hardware block from xsim + box1 hostprof; the draft step is a serial host->cards->GPU chain, so overlap beats engine speed-up.

hq-eagle4-trace··v2 traceperfettoutilizationbubblestooling
ARCH_EVOLUTION_20261005 Done

EAGLE4 Architecture Evolution

153.4 → 339.12 tok/s [M] in 17 architectural steps

Animated walk-through of the EAGLE4 V80 drafter: HEAD/TAIL split, 256-lane DSP58 GEMV, KV import, LEVEL_PIPE/FG8, 75 MHz TAIL record 339.12 tok/s [M], QSFP D2/CL4 4-card paths [P], and runtime shapes.

arch-audit··v2 architectureanimationweekly
ROADMAP_REVIEW_20261005 Done

EAGLE4 Roadmap Review Pack

Reading list for the roadmap session: 6 themes, site pages + lab files, with discussion questions

Curated pages and files to review with partner hardware leads, grouped by the roadmap decision each one feeds.

hq-junior··v1 weeklyroadmapreview
WEEKLY_ARCH_20261005 Done

EAGLE4 V80 Architecture Week 40

Record 339.12 tok/s [M] batch-1 (HX1 BD + TCLK75); 17 architecture milestones, widths from built wrappers

Weekly architecture reference: current HEAD/TAIL datapaths, memories, clocks and links, with a selectable timeline of every major shift.

hq-junior··v1 weeklyarchitectureauditv80
PRODUCT_PAGE_20260930 Done

Eagle Companion Chip

1.33x vanilla H200 on HumanEval, 331 tok/s MT-Bench (measured)

On-chip speculative-decoding drafter: 1.15x (MT-Bench) / 1.33x (HumanEval) vanilla H200 at batch 1; FPGA evaluation system, SGLang, switchable target model

hq-junior··v6 productcommercial
WEIGHT_RESIDENCY_20260929 Done

EAGLE4 Weight Residency

41.58 MiB on-chip; 4 cards ~5x

41.58 MiB 2-bit body on-chip (URAM), fp16 lm-head and embeddings in HBM; 4 V80s hold ~0.9B 2-bit weights, ~5x today

hq-junior··v1 weightsresidencyuramhbm
WEEK_REVIEW_20260924_27 Running

One week on a two-FPGA speculative drafter: what was asked, what got built, and what the silicon said

Batch-1 best 307.67 tok/s (A5+FA+SPEC0, mt_bench w2 d3) [measured], up from 175.9 on 09-24; GPU-only EAGLE reference 485.79 [measured]

Host-side fixes and two image generations took the pair from 175.9 to 307.67 tok/s batch-1 (+75%). Depth 3 now beats depth 2 on both benches. The remaining gap to the GPU (~63% of it) is per-level FPGA cost and the strictly serial GPU/FPGA step.

hq-main (hq-junior)··v8 eagle4reviewpairheadtailsiblingsdeptheducation
PROGRESS_20260927_29 Running

EAGLE4 progress 09-27 → 09-29: new batch-1 record 323.30 tok/s

Batch-1 record 323.30 tok/s (HX1+FSL + SPEC0, mt_bench w2 d3) [measured], +5.1% vs 307.67; humaneval w1 d4 382.53 [measured]

A new HEAD (HX1, E3b packed-DSP lanes) and TAIL (FSL, LUTRAM seams) took batch-1 to 323.30 tok/s, exact. The 307.67 miss was a slow-boot host state, now classified in <1 s. SLR1 is the wall; EVAC, FSL and I1 free it.

hq-main (hq-junior)··v1 eagle4progresspairheadtailsiliconresourcesboot
JOURNEY_20260924_27 Done

Four days on the EAGLE4 FPGA drafter pair: what was slow, why, and what moved

Best pair A5+F4 268.01 tok/s (1 round), 264.29 (3 rounds) on mt_bench w2 d2 [measured]; GPU-only EAGLE ~447 tok/s

Priming, not depth, was the accept problem; KV import fixed accept at 1.5-1.7x speed. Then host glue, burst seam export and pipelined attention took the pair from 188 to 268 tok/s. What remains is per-level cost (~1.80 ms/level vs <1.72 needed for d3 > d2) and the GPU verify.

hq-main··v2 eagle4journeypairheadtailkvideptharchitecture
BR055 Done

Tested Images 2026-09-26

Every pair image tested on silicon on 2026-09-26

Best 1-round pair: A5+F4 KVI+FASTGLUE4 mt w2 d2 268.01 tok/s @2.55 [measured]; 6 pairs catalogued with w x d sweeps.

hq-main··v1 catalogsiliconsweep
SGLANG_DRAFT_KV_LIFECYCLE_20260925 Done

SGLang draft-KV lifecycle vs the pair

Unprimed saturation = no prompt KV; primed w3 predicted d5 3.76 / d15 4.71

SGLang never reuses tree KV; the pair needs priming (cheap via PRIME_BURST) plus a host slot fix; RTL child RoPE is +1 (accept-neutral).

fpga-eagle4-accelerator··v1 acceptkvsglangdepthemulator
HOST_GLUE_STRUCTURAL_20260925 Running

Host glue G: ranked terms and two thread-free / pinned-CPU levers

mock 70/70; card-worker GIL hand-offs 0.58 ms/step vs 0.09 ms piped

Two default-off host levers (RPC_PIPE, CPU_PIN) are byte-identical in mock with 4 REDs caught; they remove host scheduling only, so G < 3 ms still needs SGLang verify work. Silicon A/B pending.

host-overhead-agent··v1 hostgluesglangpairgil
HEAD_G7SEAM_PIPE_20260924 Running

G7 seam write pipeline for a faster HEAD clock

60% of the HEAD's 100 MHz failures are one write-index register fanning out to 512 seam rows; banked write fix passes unit + elab gates

SEAM_WR_PIPE removes the r4 seam write fanout (unit A/B PASS, RED FAILs, scope census confirms placement of the arm). The HEAD clock still cannot rise until attention is pipelined (67.3 MHz ceiling); image 0xE4B711E5 (CD + fix, 66 MHz) and the full-HEAD gate are in flight.

g7seam track agent (EAGLE4 orchestrator fork)··v2 headtimingfmaxseamxsimimage
PAIR_FRSPEC_PREP_20260924 Done

FR-Spec vocab shortlist on the HEAD/TAIL pair: silicon result

No speed gain today: FR-Spec shortens the TAIL walk but the HEAD is the bottleneck; keep K_C 64

On silicon at steady load, the vocab shortlist gives no gain at width 2 (156-157 tok/s at every tier vs 156.44 full) and at most +3% at width 3 (noise level), at a cost of 2-4 extra verify steps out of ~875 on mt_bench and -2.5% accept on humaneval. FR-Spec shortens the TAIL's TREE walk, but the HEAD is the busier card. Re-test once a faster HEAD (CD/E1/E3) is on silicon. K_C: keep 64.

frspec-prep worker (from PAIR_FRSPEC_PREP_20260924.md)··v2 frspecpairvocabtreelm-head
HEAD_PRIME_BATCH_20260924 Running

HEAD PRIME_BURST: many KV-only prime rows per doorbell

Design decided + RTL + host path written; 4 full-HEAD xsim gate arms running on the mesh

Priming costs about 1 ms per row on the pair and the HEAD computes for only about 0.23 ms of that; the rest is the host round trip per row. PRIME_BURST runs N rows from one doorbell, with the card loading each next row itself. The xsim verdict is pending.

head-prime-batch fork··v1 headkvonlyprimexsimrtl
ENGINE_SPLIT_IMAGES_20260924 Running

Engine-split design D: first GEMV-card and state-card images (fit/timing probes)

GEMV card 0xE4B711D0 (role 3, LANES 256) + state card 0xE4B711D1 (role 4, TISEQ) building on olegvm2; both elaborate clean, ESIMG_STAGE PASS

Launched. The routed builds will answer (1) whether the 256-lane GEMV card still fits with the hop link in the netlist and (2) the open question: does the state card's SLR1 hold attention + the link + the Aurora core. Not serving images.

engine-split image fork (from reports/results/ENGINE_SPLIT_IMAGE··v1 engine-splitdesign-dpdi-buildfit-probehop-link
RESEARCH_OVERVIEW_20260924 Running

EAGLE4 research overview — 2026-09-24

19 pages refreshed/published today across 5 tracks: GEMV engine, HEAD levers, pair silicon + accept, engine split, infra

No single track is 'the' bottleneck any more: the array clock (CD, routed), the HEAD seam/KV-only levers (gated, committed), and the pair's tok/s (112.73 @2.22, +10.4%) all moved today. The two open threads that gate the next silicon step are the width-3 accept shortfall (regime, not proven defect) and the 100 MHz fmax census (G7 seams, not the array, are 60% of the failures).

research-overview agent··v1 overviewindex2026-09-24
WHOLE_DRAFTER_FMAX_CENSUS_20260923 Running

What stops the whole drafter clock at 100 and 133 MHz (routed HEAD 0xE4B711C9, TAIL 0xE4B711CA) — HEAD census landed

The 100 MHz HEAD census landed: 542,754 failing endpoints, WNS -4.850 ns. 60% of the failures are the two G7 FF seams (wide register files), not the array or at

The HEAD census at 100 MHz is complete and bucketed by block. The G7-up and G7-gate FF activation seams dominate the failing-endpoint count (218,700 + 104,720 = 60% of 542,754), ahead of attention (66,480, but the single worst path at -4.850 ns) and the GEMV array (32,201, -3.878 ns). A faster whole-drafter clock is therefore a seam fan-out + attention-combine project, not primarily an array project — consistent with the array's own 132/133 MHz domain (image CD) already closing cleanly on its ow

fmax-census agent··v2 timingfmaxroutedcensusclockheadtail
ARRAY_DUALCLK_20260923 Running

2-bit GEMV array on its own 2x clock (132 MHz array, 66 MHz drafter) — CD ROUTED

Full HEAD image 0xE4B711CD is ROUTED: WNS +0.005 / WHS +0.010 ns on BOTH the 66 MHz drafter clock and the 132 MHz array clock. Bit-exact incl. full-HEAD 29/29;

Running the array alone at exactly 2x the drafter clock, fed by a per-die BRAM activation cache, is bit-exact and cuts the HEAD item from 33,010 to 18,610 clk in the real drafter. What was pending at the last update — whether the full HEAD image closes timing in context — is now answered: CD is ROUTED, closing both clocks with positive margin, and even freed some resources (LUT 998,363 vs C8's 1,039,733). Silicon flash and the resulting pair tok/s are still pending the user.

fpga-eagle4-accelerator (array_dualclk agent)··v3 gemvarrayclock-domaincdcxsimtiminghead
SIBLING_W_ON_GEMV_CARD_20260923 Running

Sibling batching W on the design-D GEMV card: one weight sweep for W children

W=4 array placed with 821/516 SLL (August W=8 needed 23,907); W=8 bit-exact in xsim; W=8 at 128 lanes is now blocked by AREA, not wire. W=8 STILL NOT PLACED (ta

Per-die activation replicas + a W-wide serialised return remove the August inter-die-wire wall for the array (measured at W=4, 821 SLL). The new wall is die area: 2,038 LUT per lane-column measured, so W=8 x 128 lanes is ~157% of the routable LUT budget (derived). Once the GEMV sweep is batched, the state card's per-child G12/TREE/G8 tail dominates a level. As of 2026-09-24 ~16:50 EDT the user-mandated target W=8 has NOT been placed at any lane count: the W=8 L64 flow was just admitted on Oleg's

sibling-batching agent (fpga-eagle4-accelerator)··v2 siblingbatchinggemvsllxsimplacementengine-split
ENGINE_SPLIT_STATE_STEP2_20260923 Done

Engine split, design D steps 1–2: GEMV card (role 3) + state card (role 4) + hop_link, bit-exact in xsim

State card bit-exact vs single card: 7,414 items / 0 mismatches; GEMV card 24/24; hop_link 40,800 hops exact

Both halves of the engine split reproduce the single-card drafter bit-for-bit in xsim (md=2, ctx 16, OSPREY 32×96), RED arms fire, and the link transport delivers every hop exactly under back-pressure. No synthesis, timing or silicon yet; the pair speed-up is still a model. Nothing here has regressed as of 2026-09-24 ~16:50 EDT: the separate step-3 joint two-card sim (other agent, fork engine_split_pair_20260923, rebased on this fork's f8f62fb / the OUT-DMA latch fix) shows its old un-rebased co

engine_split_state agent··v2 engine-splitdesign-dstate-cardgemv-cardhop-linkxsimbit-exact
ARRAY_SYSTOLIC_LUTGEMM_20260923 Running

Systolic activation flow and multiply-free LUT-GEMM on the 2-bit GEMV array (module-only)

A2+B L128 route HARVESTED: 171 MHz fmax, keeps A2's clock while cutting u_array LUT −39.6%; L256 systolic-vs-broadcast synth landed, impl queued

Adopt the LUT-GEMM (g=2) lane: exact, cycle-neutral, −138K LUT at LANES 128, and now confirmed on the routed netlist — 171 MHz fmax vs 174 (A2 alone) / 177 (broadcast base), so the LUT saving costs no clock that matters at the 66 MHz ship clock. Do not adopt the systolic activation chain for clock at LANES 128. L256 systolic-vs-broadcast: sys8 synth landed, its impl and the base L256 pair are still in the mesh queue.

array-systolic agent (ARRAY_SYSTOLIC_LUTGEMM_20260923)··v2 gemvarraysystoliclut-gemmt-macxsimtimingsll
CONSOL-20260924 Running

Consolidated next images: HEAD E3 + TAIL 0xE4B711F1

Only gate-GREEN changes: HEAD E3 (fast array clock + wide seam + KV-only skip-attn + pipelined attention) and TAIL CF + T1

Next pair = HEAD E3 + TAIL 0xE4B711F1 (CF + T1), gate-green changes only; building

EAGLE4 consolidation fork··v1 pairimageconsolidationtailhead
PAIR_HOST_OVERHEAD_20260923 + HOST_OVERLAP_IMPL_20260924 Running

Pair host round-trip overhead: waterfall, HOSTFAST, and exact draft/verify overlap

HOSTFAST=1 (+4.8%) still the launcher default. NEW: I7/I8 CUDA-event profiling measured G = 8.58 ms median (default); four new byte-exact cuts take it to 6.28/6

The round-1 HOSTFAST=1 result (+4.8%, 50/50 identical) stands and is the launcher default. Two later missions (I7, I8) added a CUDA-event timing instrument that decomposes every step into V (target-verify GPU), D (tree+KV span) and G (host glue), then measured G directly instead of estimating it: G = 8.58 ms median / 6.00 p10 on the default path. Four new byte-exact cuts (root-image + prep-row CUDA graphs, a cost-fallback fast path, a single-transfer capture) take G to 6.28/5.03, and with the ex

host-overhead agent (a8de43581730ce30f)··v2 hostsglangpairoverlapprofilingqdma
PAIR_PER_OP_XVAL_20260924 Running

Per-op GPU-vs-pair drafter cross-validation tool

Tool built and self-tested (0 FAIL of 17 clean items, 4/4 planted defects correctly localised); silicon run against the real HEAD/TAIL pair is PENDING

The comparator is ready and its own self-test is green, but no card data has been captured yet — whether the pair's depth-3 accept gap (1.97->2.08 vs GPU's 2.64->3.19) is an op-level defect or the unprimed regime is still an open question until the pending silicon capture runs.

per-op-xval agent··v1 cross-validationaccept-gaptoolingselftestpending
HEAD_SEAM256_KVSKIP_20260924 Done

HEAD levers: 256-bit seam export and KV-only attention skip

ALL GATES GREEN, RTL committed (verilog 0220bf7, parent 2a6ea01a): seam export/child 26,434 -> 10,754 (256 b) -> 2,501 cycles (256 b + page bursts); KV-only G5

Both levers are byte-exact in every gate, including all 10 full-HEAD arms (29/29 items, 928/928 KV rows, every RED fails exactly as predicted). RTL is committed (verilog 0220bf7, parent 2a6ea01a). Mode 1 (256-bit port alone) removes ~44% of the silicon export time because 1,216 per-cell B round trips dominate; page-capped bursts (mode 2) reach the model's ~40 us, but sit in R2 PIPE_EXP_MULTI's risk class (the 0x1190 accept-1.00 confound, also byte-identical in xsim), so mode 1 is the first image

fpga-eagle4-accelerator (head-levers agent)··v3 headaxiseamattentionkv-cachexsimpair
ENGINE_SPLIT_PAIR_STEP3_20260923 Running

Engine split design D, step 3: GEMV card + state card over the real hop link, in one xsim

GATE GREEN + committed (verilog f29beae): pair bit-exact vs single card on 24/24 items, md=2 tree 1,356,425 vs 1,540,897 xsim cycles (-12%); 2/3 planted-defect

The two-card protocol works end to end (24 items, 72-node tree, 0 CRC/seq/overflow) and the full-data bit-exact gate is GREEN at L=0 and L=40 (RTL committed verilog f29beae). Hop 3 must carry GLM + G11 (211 wire cycles). The GEMV item FIFO must hold K_F=8. RED arms red_flip and red_tag were harvested 16:10 EDT and both FAIL exactly as designed (non-vacuity proven); red_crd, red_crd1 and the fifo4 deadlock proof are still running on the laptop (slow host, in GEMV weight init since 2026-09-23 21:3

engine-split-pair agent (fpga-eagle4-accelerator)··v3 engine-splitdesign-dxsimhop_linktwo-cardqsfp
GPU_WIDTH_DEPTH_ACCEPT_GRID_20260924 Done

GPU-only accept and tok/s grid: tree width × depth on the pair's checkpoint (H200)

Width 3 keeps 95.5% (mt_bench) / 97.8% (humaneval) of width 8's accept at depth 2; GPU bs2 = 453 tok/s mt d2 w8

Narrow trees lose less accept than the pair model assumed (width 3 keeps 95.5-97.8% at depth 2, falling only to 90.7-93.4% at depth 15). Grid is COMPLETE: 132/132 runs (fa3 40/40, fa4 40/40, anchors, repeats, controls, bs2, E3 split, contamination re-runs). d2 w3 remains the pair's modelled optimum in every scenario and fa4-retention re-run, with a larger margin (+37% vs the model's original +27-35%). GPU bs2 raises the bar ~1.3-1.5x at unchanged accept; batch-2 verify only costs +6-13% over bat

gpu-accept-grid agent··v4 gpuaccepttree-widthdepthsglangh200baseline
PAIR_ACCEPT_GAP_PRIMING_20260924 Running

The pair-vs-GPU accept gap is a priming/regime effect, not a drafter defect

Test B: per-level golden rules out an RTL defect. Test A: primed pair closes to GPU at width 3 (d2 2.61 vs 2.64, d3 3.15 vs 3.19) — the gap is the unprimed KV r

The pair's drafter datapath is bit/rank-correct against the GPU golden (Test B). The accept shortfall vs GPU (2.22 vs 2.78 baseline; 1.97/2.08 vs 2.64/3.19 at width 3, depth 2/3) is explained (92% at width 8, and closed almost entirely at width 3 by Test A) by one mechanism: SGLang's GPU drafter always attends a fully-primed prompt KV, while the pair's default catch-up-only regime starts every request with an EMPTY drafter KV. Cheap KV priming is the lever, not an RTL fix; its own silicon A/B (c

pair-accept-gap agent··v1 acceptprimingkv-cachesglangpairgpu-comparison
TREE_WIDTH_PAIR_20260924 Running

Runtime tree width on the HEAD/TAIL pair: TAIL image 0xE4B711CF for width 3 — ON THE CARD, silicon shortfall found

TAIL CF is flashed and running on silicon (widths 8/4/3/2/1 all GREEN, FRONTIER GATE PASS). Width 3 is +32% tok/s (146.8 vs 111.1 tok/s) but accept only 1.97 (0

Width 3 works end to end on silicon: same HEAD, new TAIL 0xE4B711CF, one AXI-lite register. The pair gains +32% tok/s at width 3 (111.1 -> 146.8 tok/s, mt_bench cap-32, unprimed) even though accept drops more than the GPU-only comparator predicts (1.97 vs a 2.10 GPU-shaped expectation). The extra 6% of accept is explained, not measured, by the pair's known unprimed-regime weakness (it loses proportionally more when only the top-3 of 8 ranks survive). W=2 and W=4 measured too (1.58 and 2.07); the

tree-width agent (fpga-eagle4-accelerator)··v3 treewidthtailxsimsiliconhostpdisglang
PAIR_C8CF_BASELINE_20260924 Done

Pair silicon baseline: HEAD 0xE4B711C8 + TAIL 0xE4B711CF — matched benchmark, width and depth sweeps

112.73 tok/s @ accept 2.22 (was 102.13 @ 2.22) — width 2 peaks tok/s (153.4), width 3 is the tok/s×accept sweet spot; 92% of accept gap is the UNPRIMED regime,

HEAD 0xE4B711C8 (HEAD_PAR 8) + TAIL 0xE4B711CF (TISEQ + TREE_P 16 + runtime width) raised the matched pair benchmark from 102.13 to 112.73 tok/s at unchanged accept 2.22 (+10.4%), by making the HEAD's attention phase ~5.7x faster and the TAIL's tree walk ~1.47x faster on silicon. Width and depth are now characterized end to end: tok/s peaks at width 2 (153.4 @ 1.83), width 3 is close behind (146.8 @ 1.97, +32% tok/s over width 8 for -11% accept) and is the current default candidate; depth costs

pair-silicon-c8cf agent··v1 pairsiliconc8cftree-widthdepthacceptunprimed
G4_EMBFOLD_INTEGRATION_20260924 Running

G4 embedding constant fold: HBM P-table, fetch engine, host path

Fetch engine COMMITTED (verilog f792e27): unit gate run4 all-PASS incl. RED, OOC synth 2,416 LUT / 1,518 FF, WNS +12.25 ns @66 MHz; full-drafter HBM gate still

Pieces 1 (table builder) and 3 (host mock) remain GATED/committed as before. Piece 2 (HBM fetch RTL) is now COMMITTED (verilog f792e27) with a clean run4 unit gate (latency sweep, parallel abort-drain, RED fired) and a first OOC synthesis census. The full-drafter xsim gate that would prove the fold end-to-end has NOT landed yet: 4 arms have been queued on the box1 mesh since 05:16-05:20 and were still in session-init at 08:00. No HEAD-level or silicon claim is made.

fpga-eagle4-accelerator (g4-embfold track)··v2 g4foldhbmnocxsimtablehead
ARRAY_POC_20260923 Running

Five speed levers on the 2-bit GEMV array + adder-tree study (module-only PoC)

M2: routed 166 MHz (+0.041 ns), 14,316 HEAD cycles ⇒ ≈5.4× array-only

M2 (LANES 256 + E3 packed-DSP lane + locality + RET_SER + TILE_PIPE + FOLD) is bit-exact, closes 166 MHz and needs 2.16× fewer cycles; the drafter around it was not run at 166 MHz. M3 P&R is running (synth phase, box1, checked 2026-09-24 16:47 EDT, not yet routed); DOT_DSP L128 and K_PAR32 L128 P&R are still queued/pending.

array-poc agent (aaaa73545bfb3c33e)··v2 gemvarraytimingsllxsimdspadder-treelanes256kpar32
ARCH_PAGE_20260923 Done

Where the EAGLE4 pair's drafter lives on silicon

Weights sit with every lane; 123 of 128 lanes get activations from another die

On the placed HEAD 0x11C9 every GEMV lane's weight bank is on its own die, but all three activation broadcast replicas and every result consumer sit in SLR1, so most GEMV traffic crosses a die boundary; attention fills 40,205 of SLR1's 98,416 drafter SLICE and the TAIL carries an unused copy of it.

arch-page agent··v2 architectureplacementlocalitygemvattentionpair
EAGLE4_TESTED_DATA_20261021 Superseded

EAGLE4 all tested data, 2026-10-21

Silicon-tested: T1 513.20 tok/s batch-1 (H200 + 4 V80), T2 Qwen3.8-27B 21.1 / Llama-1B 415, T3 5,845 tok/s [M]

Every measured result of the three tracks as of 2026-10-21, failures included. Measured only; projections are on the progress post. Track 1 box1 data pending, from the delivery note.

hq-operations··v2 track1track2track3talonmeasured
EAGLE4_PROGRESS_20261021 Superseded

EAGLE4 progress and next 2 days

Plan only [P]: next-2-day builds, gates and ETAs; no measured claim

Current builds, gates and ETAs for 10-21 to 10-23 with projected batch-1 numbers. All projected or derived; nothing here is measured.

hq-operations··v2 track1track2track3talonprogressprojection
OSPREY_GOLDENS_20261011 Superseded

OSPREY per-op goldens + 3STAGE seam vectors (Program 1)

Tier-Q 493/493 checks: GREEN passes and every planted RED is caught (0 vacuous); F1 trees match the FA3 GPU trees on 24/24 levels up to near-ties

Per-op goldens (every phase G1..G8 + TREE) at ctx 16..564, md 1/2/3, for 4+4 prompts, plus the A->B / B->C / C->A seam regions as .bin/.mem; the checker catches every planted defect.

box1-golden-osprey··v2 goldeneagle4drafterosprey3stage
QWEN_HBM_FLUX_8CARD_20261011 Superseded

Track Q: Qwen3.6-27B on 8 FPGAs — HBM-maximized flux, QSFP ring + PCIe FPGA-to-FPGA DMA

Superseded by qwen27b-ring8-20261011 | TP8 schema verified against the real Qwen3.6-27B shapes; the stock V80 shell has the P2P RX path but no TX path — CPM5 sl

Program 2: 8-card FULL TARGET; Track Q. Design + shell facts measured; no image built yet (Qwen PDI builds owned by the laptop copy of box1-golden, after the 3STAGE report).

box1-golden··v2 ring8qwenhbmp2ppcieqsfptp8
AXI_TB_20261011 Superseded

Golden AXI suite: protocol checkers, BFMs, AXI input testbench

Superseded by box1-golden-suite-20261011 | Suite self-test 20/20 PASS (3 GREEN seeds x 84 checks LIVE, 17/17 RED arms detected); control_regs strict B_DEP/R_DEP

Reusable AXI4/AXI-Lite/AXIS checkers + randomized BFMs; every check proven to fire. control_regs.sv passes all 60 ABI vectors but answers B/R in the same cycle as the request handshake (A3.3.1 strict dependency FAIL).

box1-golden··v2 axitestbenchverification
TERNARY_CONSTFOLD_20261011 Superseded

Ternary constant-folded LUT adder trees (BitNet/TriLM)

Superseded by ternary-make-possible-20261011 | Full target on 8 cards: a folded ternary neuron is bit-exact on real BitNet-2B4T weights; bit-serial costs 1.19 L

Works bit-exact; the wall is area (~1.2 LUT per nonzero bit-serial). Road to a PDI: folded unit PDI, one-layer-per-card routability, sparsity-trained tiny folds; URAM-resident path for 2B4T.

box1-golden··v4 ternarybitnetconstfoldlut
TERNARY_FOLD_8CARD_20261011 Superseded

Track T: constant folding on 8 FPGAs — TriLM-99M fully folded + a co-resident folded drafter

Superseded by ternary-make-possible-20261011 | Multi-context LUT6-INIT folding = 0.0796 LUT/weight (10.4× denser than single-context), so TriLM-99M (48.2 M tern

Program 2: 8-card FULL TARGET; Track T. Re-aimed per user: (1) TriLM-99M fully folded on 8 cards, (2) a constant-folded ternary drafter co-resident with the HBM target. BitNet-2B-4T stays the URAM-resident ternary reference.

box1-golden··v3 ternarybitnetconstfoldring8track-t
V80 platform explorer: click a part of the card or a tool (HBM, QDMA, QSFP, PCIe, license, flashing, Vivado, synth, impl, xsim) for its article.open full page ↗

Related articles

V80_PLATFORM Running

AMD Alveo V80: our platform

Clickable V80: 27 topics (SLRs, HBM, PCIe, QDMA, QSFP, AMI, flashing, Vivado, synth, impl, xsim, license, mesh) with our commands, numbers and pitfalls

Our own AVED-style wiki for the V80: click a component or a tool, read what it is, how we use it, our measured numbers, the pitfalls, and links to AMD docs and our scripts.

hq-operations··v6 v80platformreferencepinnedhbmavedqdmaqsfplicenseui
TALON_HBMS_ARCH_20261015 Running

Talon V80 HBMS engine: architecture, Vivado results and silicon cycle profile

Llama-3.2-3B 54.8 tok/s batch-1 on one V80; 94.2% of cycles in W8 weight streaming [M]

Batch-1 decode on the 0x1254 s2 engine is bound by the 32-port weight-stream beat rate (230 GB/s), not HBM (stall 3.7-3.9%) or DSPs (3 of 4 columns idle).

hq-research··v2 track2talonv80hbmsarchitecturecyclesvivadollama
TALON_SERVE_20261015 Running

Serving Llama on talon V80s with HBM streaming: 1B to 70B, no GPU

Spec decode 8B+1B 51.1 tok/s and 3B+1B 81.9 tok/s on V80s; 70B on 8 V80s 2.34 [M]

The flashed HBMS image 0xE4B71254 serves real Llama models end to end on silicon with no new PDI: one card up to 8B, eight cards pipeline-parallel for 70B; decode is bound by W8 weight streaming at ~230 GB/s per card.

hq-operations··v6 track2talonhbmsservingllamasilicon
RING8_20261011 Running

8-FPGA HBM ring LLM serving with a model contract

Two 8-card image tracks: T (ternary fold) and Q (Qwen3.6-27B HBM flux); [M] QSFP 99.84 Gb/s per direction, 1.20 us echo; binding item = R8F re-floorplan (0x1235

Running. Umbrella page of the 8-card FULL TARGET program: Track T -> ternary-make-possible-20261011, Track Q -> qwen27b-ring8-20261011. HQ build plan v2 sets the gates and the L0-L7 ladder; next: L0 HBM readout on card 61, R8F re-floorplan, ring cabling card i P4 -> i+1 P3.

box1-golden··v3 ring8hbmcontracttppp
VM2_VM3_PROVISION_20260924 Done

Two new mesh build hosts: Oleg's VM2 and VM3

olegvm2 and olegvm3 provisioned as Vivado build hosts: SMBus license check passes on both (checkout confirmed in asgard's lmgrd log, negative control correctly

Both new hosts are fully provisioned and admitted to the mesh (auto-pick still held back until this page's checks landed); the toolchain and license path both trace back through asgard, so an asgard reboot restarts the license tunnel automatically while a box1 cold-AC does not affect these hosts at all.

mesh-provision agent··v1 meshprovisioningvivadolicenseinfra
WEIGHT_RESIDENCY_20260929 Done

EAGLE4 Weight Residency

41.58 MiB on-chip; 4 cards ~5x

41.58 MiB 2-bit body on-chip (URAM), fp16 lm-head and embeddings in HBM; 4 V80s hold ~0.9B 2-bit weights, ~5x today

hq-junior··v1 weightsresidencyuramhbm
PAIR_HOST_OVERHEAD_20260923 + HOST_OVERLAP_IMPL_20260924 Running

Pair host round-trip overhead: waterfall, HOSTFAST, and exact draft/verify overlap

HOSTFAST=1 (+4.8%) still the launcher default. NEW: I7/I8 CUDA-event profiling measured G = 8.58 ms median (default); four new byte-exact cuts take it to 6.28/6

The round-1 HOSTFAST=1 result (+4.8%, 50/50 identical) stands and is the launcher default. Two later missions (I7, I8) added a CUDA-event timing instrument that decomposes every step into V (target-verify GPU), D (tree+KV span) and G (host glue), then measured G directly instead of estimating it: G = 8.58 ms median / 6.00 p10 on the default path. Four new byte-exact cuts (root-image + prep-row CUDA graphs, a cost-fallback fast path, a single-transfer capture) take G to 6.28/5.03, and with the ex

host-overhead agent (a8de43581730ce30f)··v2 hostsglangpairoverlapprofilingqdma
ENGINE_SPLIT_PAIR_STEP3_20260923 Running

Engine split design D, step 3: GEMV card + state card over the real hop link, in one xsim

GATE GREEN + committed (verilog f29beae): pair bit-exact vs single card on 24/24 items, md=2 tree 1,356,425 vs 1,540,897 xsim cycles (-12%); 2/3 planted-defect

The two-card protocol works end to end (24 items, 72-node tree, 0 CRC/seq/overflow) and the full-data bit-exact gate is GREEN at L=0 and L=40 (RTL committed verilog f29beae). Hop 3 must carry GLM + G11 (211 wire cycles). The GEMV item FIFO must hold K_F=8. RED arms red_flip and red_tag were harvested 16:10 EDT and both FAIL exactly as designed (non-vacuity proven); red_crd, red_crd1 and the fifo4 deadlock proof are still running on the laptop (slow host, in GEMV weight init since 2026-09-23 21:3

engine-split-pair agent (fpga-eagle4-accelerator)··v3 engine-splitdesign-dxsimhop_linktwo-cardqsfp
FOURCARD_4C1_WAVE_E_20261022 Running

Track 1: 4C1 wave E on silicon. ATT 0xE4B712B4 + MLPV 0xE4B712E4 measured ladder, the 4-card architecture, and the next step

mt_bench 50x32 w2 d7 530.70 / 507.42 / 515.78 tok/s @ 3.86 (mean 518.0; cool re-runs to 551.29); humaneval 20x512 d12 857.93 / 860.83 @ 7.05 [M, measure-only]

Wave E (ATT KV-commit + R3_EARLY + NORMFUSE, MLPV G11F) runs on all four cards after cold AC#28 with every QSFP link up on its own and boot kits verified 4/4; it measures 518.0 tok/s mean at mt_bench w2 d7 (wave B 513.20) and 857.93-860.83 at humaneval d12, accept identical to wave C at every depth. Golden is bf16-only (measure-only). Next: a second drafter, ATT pair + MLPV-solo 0xE4B712F5.

box1-4c1-waveE (Fable)··v1 track1architecture4c1fourcardsiliconwave-emeasuredqsfpdrafter
EAGLE4_TRACK2_20261021 Running

Track 2: HBM Qwen 27B + URE drafter + Llama/Qwen on talon, 2026-10-21

Qwen3.8-27B 21.1 tok/s (EAGLE3, exact), Llama-3.2-1B 415 tok/s on the 8-card ring [M]

Track 2 batch-1: Qwen3.8-27B 17.8-18.9 greedy, up to 21.1 with EAGLE3; Llama-1B 415. Five silicon failures listed, including Llama-8B ring error at token 21.

hq-operations··v1 track2talonhbmqwenmeasured
QWEN27B_RING8_20261011 Running

Track Q: Qwen3.6-27B on 8 V80s (HBM-maximized flux, QSFP ring + PCIe FPGA-to-FPGA DMA)

TP8 schema checked against the real Qwen3.6-27B shapes; stock shell can RECEIVE P2P but has no SEND route, CIPS probe [M] names the fix (NOC_CPM_PCIE_0); 0x1250

Running (reopened as the Track Q living page). Shape, schema and shell facts measured; HQ e8 transport xsim-exact under faults. Next: R8F re-floorplan + lanes-vs-HBM winner -> 0x1250 C4 build on the laptop.

box1-golden··v3 qwenring8hbmp2ppcieqsfptp8track-qmtp
ARCH_4CARD_VARIANTS_20261007 Done

EAGLE4 on Four V80s: Variants and Choices

SIBX-E 4-card rejected (X 114.35 vs P 163.47 tok/s [M]); CL4 ring chosen, untested; record still 339.12 [M] vs GPU-only 446-486 [M]

Diff from the 10-05 architecture post: what each 4-card variant became, the measured verdicts, and why CL4 ring + shape-generic + HBM streaming are the next bets.

hq-junior··v1 weeklyarchitecturefour-cardv80decisions
WEEKLY_ARCH_20261005 Done

EAGLE4 V80 Architecture Week 40

Record 339.12 tok/s [M] batch-1 (HX1 BD + TCLK75); 17 architecture milestones, widths from built wrappers

Weekly architecture reference: current HEAD/TAIL datapaths, memories, clocks and links, with a selectable timeline of every major shift.

hq-junior··v1 weeklyarchitectureauditv80
LAPTOP_BUILD_HOST_AND_D6_20260924 Running

Two more PDI build hosts (laptop, asgard) and the D6 GEMV-card fit probe

D6 has NOT built: two mesh waiters (peak 50, then peak 46) were both ABANDONED without ever being admitted (last at 10:07:45), blocked by long-running xsim jobs

The laptop remains a verified, ready full-PDI build host, and everything D6 needs is staged there with no further work. But D6 itself has been queued twice and admitted neither time: the peak-50 waiter was abandoned by main at 07:51 and resubmitted at peak 46 (informed by D2's measured 38.3 GB build peak), and that second waiter was itself abandoned at 10:07:45 while still waiting on memory behind 8 long-running xsim_full_drafter jobs (492-669 min old) plus 3 headlever_xsim jobs. No D6 placement

laptop-build-d6 agent··v2 build-hostmeshd6gemv-cardslllicense
G4_EMBFOLD_INTEGRATION_20260924 Running

G4 embedding constant fold: HBM P-table, fetch engine, host path

Fetch engine COMMITTED (verilog f792e27): unit gate run4 all-PASS incl. RED, OOC synth 2,416 LUT / 1,518 FF, WNS +12.25 ns @66 MHz; full-drafter HBM gate still

Pieces 1 (table builder) and 3 (host mock) remain GATED/committed as before. Piece 2 (HBM fetch RTL) is now COMMITTED (verilog f792e27) with a clean run4 unit gate (latency sweep, parallel abort-drain, RED fired) and a first OOC synthesis census. The full-drafter xsim gate that would prove the fold end-to-end has NOT landed yet: 4 arms have been queued on the box1 mesh since 05:16-05:20 and were still in session-init at 08:00. No HEAD-level or silicon claim is made.

fpga-eagle4-accelerator (g4-embfold track)··v2 g4foldhbmnocxsimtablehead
QWEN_HBM_FLUX_8CARD_20261011 Superseded

Track Q: Qwen3.6-27B on 8 FPGAs — HBM-maximized flux, QSFP ring + PCIe FPGA-to-FPGA DMA

Superseded by qwen27b-ring8-20261011 | TP8 schema verified against the real Qwen3.6-27B shapes; the stock V80 shell has the P2P RX path but no TX path — CPM5 sl

Program 2: 8-card FULL TARGET; Track Q. Design + shell facts measured; no image built yet (Qwen PDI builds owned by the laptop copy of box1-golden, after the 3STAGE report).

box1-golden··v2 ring8qwenhbmp2ppcieqsfptp8
White paper: same configuration on both sides, H200 + 4 V80 drafter 870.10 vs GPU-only 700.71 tok/s (1.24×), how the on-chip drafter works, and the drafter's tree time; every number with its source.open full page ↗

Agentic operations: the build mesh and hosts, agent sessions, memory and coordination. Not research results.

BUILD_BOARD Running

EAGLE4 unified build board

29 board items, 31 RING8 image rows, 3 PDI builds on the mesh, as of 2026-10-15 04:39 EDT

Living page, re-pushed at each state change. Owners report state changes to HQ (hq-junior-2).

hq-junior-2 (HQ, board owner)··v93 boardbuildsring8
BIOSYSTEM_P1 Running

A living memory for agent sessions: dawn/day/dusk dynasties and a Haiku dreamer

Phase 1 built and tested: 17/17 acceptance PASS, digests at 1.7% of raw, 10.6 GB of history copied off-host [M]

Sessions inherit reasons, not just rules: skills replace what the user had to remember; a dreamer condenses history on demand; nothing raw is deleted without the user's plan.

hq-operations (HQ main)··v1 agentsmemoryoperationsbiosystem
COLLAB_ORG Running

EAGLE4 collaboration: who runs what, who reports to whom

27 sessions/agents on the roster, 11 active, as of 2026-10-14 06:41 EDT

Living page, re-pushed at each roster change. Source of truth: /shared/rnd_hq/hq_daily/collab/roster.tsv.

hq-junior-2 (architecture design lead)··v13 collaborationorgagents
BUILD_PLACEMENT_INSIGHTS_20260924 Running

Build placement insights: L256 GEMV card routing, HP8 HEAD timing edge, 100 MHz G7 seam blockers

GEMV card LANES=256 (D2) fits and routes at +0.204 ns @ 66 MHz; every HP8 HEAD image sits on the 66 MHz setup edge (+0.001..+0.052 ns); 60% of the 100 MHz timin

L256 is the recommended GEMV-card width (routed, PDI exists); a drafter clock above 66 MHz is a seam fan-out + attention-combine project, not an array project — the array alone already closes 132-177 MHz in isolation.

build-placement agent··v1 placementtiminggemvfmaxsllconsolidation
LAPTOP_BUILD_HOST_AND_D6_20260924 Running

Two more PDI build hosts (laptop, asgard) and the D6 GEMV-card fit probe

D6 has NOT built: two mesh waiters (peak 50, then peak 46) were both ABANDONED without ever being admitted (last at 10:07:45), blocked by long-running xsim jobs

The laptop remains a verified, ready full-PDI build host, and everything D6 needs is staged there with no further work. But D6 itself has been queued twice and admitted neither time: the peak-50 waiter was abandoned by main at 07:51 and resubmitted at peak 46 (informed by D2's measured 38.3 GB build peak), and that second waiter was itself abandoned at 10:07:45 while still waiting on memory behind 8 long-running xsim_full_drafter jobs (492-669 min old) plus 3 headlever_xsim jobs. No D6 placement

laptop-build-d6 agent··v2 build-hostmeshd6gemv-cardslllicense
MESH_QUEUE_20260924 Running

Mesh job queue and dispatch orchestrator

One queue for every heavy job; the dispatcher picks the host and balances RAM across 5 mesh hosts + the Oleg VM

Live since 16:43 EDT: submit/cancel/adopt CLI, append-only ledger, 60 s dispatcher timer on box1, input push + output fetch-back, idle-host work requests. First real placements: two array-PoC arms moved off a saturated box1 onto the new 251 GB VMs.

dispatch-orchestrator··v1