EAGLEEAGLE CORP
Download HTMLDownload PDFOpen PDF
Calculations and hypothesis (everything on this page that is not a direct measurement)
  • Record vs GPU-only: 961.43 / 841.27 = 1.143 (+14.3 %), best FPGA setting vs best GPU-only setting; at matched depth d11 fa3: 961.43 / 825.91 = 1.164 [derived].
  • Meeting head-to-head: 870.10 / 700.71 = 1.242; same-day pooled 1.239 [derived, SOURCES.md].
  • Step split: rest of the step = step − tree path (7.187 − 2.010 = 5.176 ms); the GPU verify 3.452 ms was measured once at d7 on the 4-card set and is ASSUMED equal at d11 on 7 cards, so the host share 1.724 ms is derived from that assumption.
  • Ceiling if the card tree took zero time: step ≥ 5.686 ms ⇒ ≤ 1,215 tok/s [projection, SCALING §Calculations].
  • COMBO3 card deltas on top of 0x13BE (no census): G5_P2W −157.5 µs, R3_EARLY −112.3 µs per d11 tree [projection, SCALING §3].
EAGLEEAGLE CORP
01 / 12 · EAGLE4 · 2026-11-18

A 2-bit drafter on V80 FPGAs, beside the H200 that verifies it.

961.43 tok/sLlama-3.2-3B, one user, 7 V80 + 1 H200 (talon2), humaneval 80 × 256, accept 6.91
841.27 tok/sthe same H200 drafting for itself (best GPU-only setting, depth 9, fa3)
80 / 80replies identical text to the recorded reference, 3 of 3 rounds

measured talon2, 2026-11-16: COMBO3 0xE4B713D6 × 4 + M1361 0xE4B71361 × 2, W2 d11, rounds 960.23 / 960.97 / 963.10, texts 1b0779cf = record · DRAFTER_SPEED_RESEARCH_20261107/ATT_COMBO_SILICON_20261116.md (COMBO3 serve rows) · GPU-only: KC_RAISE_PLAN_7CARD_20261116.md:131; BEST_VS_BEST_GPU_20261112.md:4.

02 / 12 · The demo shown on 2026-11-10

Live chat: the same prompt through FPGA drafting, GPU + EAGLE, or GPU only.

box1 demo, auto-run on 4 FPGA cards

demo.yshicarrick.com runs one H200 with four V80 drafter cards (box1). A mode switch picks who drafts; auto-run loops benchmark prompts and keeps a running tally.

1,049.6 tok/sFPGA mode, auto-run best set, 5 prompts, 1,280 tokens, accept 9.08
973.32 tok/sa manual FPGA chat reply, 256 tokens, accept 7.529
269.3 tok/sQwen3.8-27B on the same cards, 10 prompts, accept 7.74

measured box1, 2026-11-10, engine E2b + MLPV 0x1342, w2 d11, cap 256; token-weighted means · DEMO_AUTORUN_BOX1_20261110.md:40-41 (Llama), :87 (Qwen); screenshot DEMO_AUTORUN_BOX1_20261110_shots/01_running.png (md5 e8fe8095…).

03 / 12 · The demo, same configuration

Same H200, same prompts, same tree: the FPGA-drafted run was faster.

demo page, desktop
870.10 tok/sH200 + 4 V80 drafting, accept 7.12
700.71 tok/sH200 alone (GPU-only EAGLE), accept 6.92
870.06 / 700.47same-day re-measure 2026-11-11, both arms

measured talon2, Llama-3.2-3B, humaneval 80 × 256, batch 1, greedy, depth 15 width 2, 20 verify tokens, fa4, SGLang 0.5.16; FPGA arm ATT 0xE4B713F9 × 2 + MLPV 0x1342 × 2 · FPGA_VS_H200_20261110/SOURCES.md (Results + same-day tables; logs md5 d75ab93e…, 86239804…). Screenshot SITE_DESIGN_AUDIT_20261110/demo_1280_light.png.

04 / 12 · Today's record (7 cards)

Seven cards: 961.43 tok/s at accept 6.91, every reply identical to the reference.

roundtok/sacceptverify stepstexts (80)
t1960.236.912,8391b0779cf = record
t2960.976.912,8391b0779cf = record
t3963.106.912,8391b0779cf = record
mean961.436.91
841.27 tok/sGPU-only best on the same H200 (d9 fa3); d11 fa3 825.91
2.878 mscard time per full d11 tree on the parent set (0x13BE × 4 + M1361 × 2)
7.187 msper verify step on the record set (2.010 ms of it waits on the cards)

measured talon2 2026-11-16, humaneval 80 × 256, W2 d11 fa3 · ATT_COMBO_SILICON_20261116.md (COMBO3 serve rows t1-t3) · KC_RAISE_PLAN_7CARD_20261116.md:131 (GPU-only) · SCALING_4C_VS_7C_PHASE_RANK_20261116.md §3 (tree), §4 (step). The ratio to GPU-only is in the calculations bar.

05 / 12 · Architecture (the Architecture Book)

One drafter split over two card roles: ATT does attention, MLPV does the MLP, vocabulary and tree.

Architecture Book block diagram, F3t 0xE4B713F9

Block diagram from the Architecture Book for its newest Track-1 image, F3t 0xE4B713F9 (4 cards): solid blocks are built, dashed are not in this image, gold = what this image added. Live, clickable book (66 images): hardware.research.yshicarrick.com/p/eagle4-arch-book/.

measured rendered 2026-11-18 from /p/eagle4-arch-book/ v5 (arch_book.json sha256 8acbaeda…), image "F3t G5_OVL ATT (0x13F9)". The book has no entry yet for the 7-card images on the next slide.

06 / 12 · Architecture of the record images

Two images, flashed on six cards: four ATT ranks share every matrix, two MLPV cards finish each node.

ATT COMBO3 0xE4B713D6 (4 cards)MLPV M1361 0xE4B71361 (2 cards)
roleG1 embedding + norm, G4 QKV, G5 attention, G6 o-proj with a 4-rank all-reduce over QSFP, G7 gate/upG9 down, G11 norm, GLM low-rank, G12 vocabulary scan + TREE top-K, G8 exact fp16 rescore
levers in this image0x13BC + G5_P2W 8 + R3_EARLY 1 + KV1H2C 1 + HANDOFF_CUT 1MLPV_G8_OWN 1 (G8 per d11 tree 551.83 → 237.71 µs vs the 4-card MLPV)
GEMV array512 lanes × 16 × 2, 66 MHz, LUT-GEMM 2-bit lanes256 lanes × 16 × 2, 132 MHz, DSP lanes
resources (one card)0x13BE parent: SLICE 81.85 % (die 0 98.12 %), DSP58 5,267 (48.6 %), URAM 744 (38.7 %), BRAM 47.7 %SLICE 89.6 %, DSP58 7,504 (69.2 %), URAM 512 (26.6 %), BRAM 27.7 %
timing (final, true)setup +0.044 ns / hold +0.010 ns, 0 failing endpointsmargin rescue e24deb3f of the same netlist
FINAL PDI md5d2e737d7c7134cb0d317345a6e405b1be24deb3f6d1de5e5b41964eecc1be7e9

measured ADJUDICATION_0x13D6_0x13DB_20261116.md:11-14 (levers, PDI, timing) · ARCH_RETHINK_7CARD_20261117.md §1.3 (resources; no utilization report exists for 0x13D6 itself, so its parent 0x13BE is shown), §1.3 "Array today" · SCALING_4C_VS_7C_PHASE_RANK_20261116.md §3 row 1 (G8) · scripts/4c1/c4_regmap.py EXPECT_PDI_MD5 (M1361).

07 / 12 · How it is set up: talon2

One server: seven V80 cards (six used) and one H200, the cards joined by QSFP cables.

talon2 7-card system

measured 13 point-to-point QSFP cables, each 4 × 25.78 Gb/s Aurora 64B/66B · SEVENCARD_BOARD_20261026/TALON2_QSFP_MAP_20261110.md:24-50 · ARCH_RETHINK_7CARD_20261117.md §1.1-1.2 (card 99 is not opened: allow-list, 4-rank hypercube, no MLPV cable).

08 / 12 · How one draft tree is made

The host sends one root; the cards build the tree level by level; the H200 verifies it in one pass.

per-tree flow

measured SCALING_4C_VS_7C_PHASE_RANK_20261116.md §3 (one full d11 tree = 21 items; ATT / MLPV phase split) · ARCH_RETHINK_7CARD_20261117.md §1.1 (stage cables, all-reduce) · FPGA_VS_H200_20261110/SOURCES.md (exec-mode 1).

09 / 12 · Where the card time goes

Per-engine time for one full depth-11 tree: attention and QKV lead on ATT, the vocabulary scan on MLPV.

cardphase4 cards (µs)7 cards (µs)
ATTG5 attention841.34610.48
ATTG4 QKV630.03452.12
ATTG7G (+ serial R3 + pre-MLP norm)561.34444.48
ATTseam / link (EMBG + EXPORT + INGEST)208.60262.71
ATTG6 o-proj + TP all-reduce195.07218.14
ATTG7U302.43177.22
ATTG1 (+R1)136.94145.09
ATTG7P / SiLU80.7442.32
MLPVTREE (G12 scan + top-K merge)701.70707.61
MLPVG9 down268.89270.45
MLPVINGEST + XCH + LVL_SEL/OUT243.25252.08
MLPVG8 exact rescore551.83237.71
MLPVG11148.05148.27
MLPVGLM73.4971.91
ATT busy2,956.492,352.56
MLPV busy1,995.221,695.03
tree issue → done, p503,384.052,877.70

measured µs per full d11 tree (21 items) from on-card counters; 4 cards = F3t 0xE4B713F9 × 2 + MLPV 0x1342 × 2, 7 cards = 0x13BE × 4 + M1361 × 2 (COMBO3's parent; no census exists for COMBO3) · SCALING_4C_VS_7C_PHASE_RANK_20261116.md §3 table (rows 1-15, totals).

10 / 12 · Where one verify step goes

Of the 7.187 ms step, the host waits on the cards for 1.501 ms; the GPU verify and host work are the rest.

term (ms per verify step)4 cards (883.12 tok/s)7 cards, record (961.43)
step7.8127.187
tree path = host_ahead + join_wait2.4552.010
· host work under the card (host_ahead)0.5120.510
· host blocked on the card reply (join_wait)1.9431.501
rest of the step (GPU verify + host)5.3585.176
GPU verify forward (cuda events, d7, 4-card F3t)3.452not measured on 7 cards

measured d11 fa3 w2, humaneval 80 × 256; steps 2,843 vs 2,839, accept 6.90 vs 6.91 · SCALING_4C_VS_7C_PHASE_RANK_20261116.md §4 · GPU verify: BATCH_BUBBLES_20261109/INTERLEAVE_BATCHVERIFY_SILICON_20261109.md:76. The split of the remainder is in the calculations bar.

11 / 12 · The run configuration

Everything needed to reproduce the record run.

target / drafterLlama-3.2-3B-Instruct (fp16) on the H200 NVL · EAGLE4 2-bit drafter bitnet_721 on the cards
workloadhumaneval, 80 prompts × 256 new tokens, batch 1, greedy
treewidth 2, depth 11, 20 verify tokens (VTOK 20), K_C 64 candidates, tree-stop theta −2.5
GPU sideSGLang, draft attention fa3, FASTGLUE 4, CUDAPACK 1
host leversHX_MMIO7 / HX_O7 / HX_S7 (fewer per-card MMIO round trips), KV1H2C on
cardsATT ranks 17 / 3d / bd / e1 = COMBO3 0xE4B713D6; MLPV 50 / 63 = M1361 0xE4B71361; drafter clock 66.001 MHz
checks before the timed roundsidentity 6 / 6 (build-id registers), golden fingerprint 51524c4f PASS, exact-tooth PASS × 2, zero KV-reader aborts

measured DW_SWEEP_7CARD_COMBO3_20261116.md:8-10 (bench line) · ATT_COMBO_SILICON_20261116.md (14:24 bring-up, 14:25:12 golden, serve rows) · SCALING_4C_VS_7C_PHASE_RANK_20261116.md:11 (clock) · FPGA_VS_H200_20261110/SOURCES.md (target, drafter weights).

12 / 12 · Sources and downloads

Every report behind these slides is in the FPGA archive.

All paths are under reports/results/ (DRAFTER_SPEED_RESEARCH_20261107/ for the record, scaling and rethink reports).