fpga-archive-reports-process · EAGLE4 FPGA archive

Source docs/Q_AND_A_20261105.md · source md5 777c3437f9ba4e9c3044bafd8652b69d · archived 2026-11-18

Download PDFOpen PDFRaw markdown

Q&A — decisions waiting on Carrick · 2026-11-05 20:40 EST

Purpose: this file holds the decisions that will eventually be needed — but they are deferred. Before ruling on any of them the session must be able to state the architecture and the development cycle correctly, which is what ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/LEARNING_BRIEF_20261105.md is for. Answer that brief first; then these six become answerable in one pass. Editing a DECISION: line is still how a ruling is made.

Cold-start: the state in ten lines

# fact evidence
1 The programme runs a 2-bit EAGLE-4 drafter on Alveo V80s; an H200 verifies the tree. Target: beat H200-only inference on the same host. ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/OPERATOR_CONTEXT_20261105.md §1
2 4C1 record parity is real: Llama 728.96 / 731.40 / 730.02 tok/s @ accept 5.88, Qwen 236–240 @ 7.14. ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/reports/results/SEVENCARD_BOARD_20261026/SIBW_BUILD_WATCH_20261026.md tail
3 The array loses end to end at bs 1: GPU-only 826.58 vs FPGA 723.09, and 1.280 vs 1.409 J/token. It wins only on drafter forward, 0.177 vs 0.37 ms (2.1×). ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/reports/results/TALON2_POWER_ECON_20261104.md, ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/reports/results/FOURCARD_DESIGN_20261016/GPU_VS_FPGA_DRAFTER_20261019.md
4 Correctness is closed. The 0x133C G5 hang was an MBUFGCE leaf-divider startup skew; two clock-only fixes exist — fix C (0x133F) and fix B (0x1340). Keep B: C closes hold only on placement luck. ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/reports/results/SEVENCARD_BOARD_20261026/RCA_0x133C_G5_HANG_20261103.md
5 L512 was never a silicon wall — it was a floorplan (hub in SLR1, consumers in SLR0). E2b 0xE4B713E2 closes at +0.043/+0.012, PDI 2540c725…, quarantined. ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/reports/results/QWEN27B_DRAFTER_20261020/LUTGEMM_ATT_INTEGRATION_D9_20261102.md §7d
6 Blocker is visibility, not engineering: asgard's LAN has been down since 11-05 07:47Z, so E0's own AMC firmware and its gate rows cannot be read. Q1 below
7 Open build: 0xE4B713EB (same netlist, E0 floorplan, no BalancedSLR) placed at +0.124 / −0.436, routing on buildx03. Q1, Q3
8 Truth hierarchy: HQ → operator pack → evidence set → domain briefs → scripts → site. Read in that order. ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/HQ_20261105.md
9 Do not read NOW.md (content ~09-14), ITINERARY.md (07-08), ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/reports/PDI_BUILD_INDEX.md (07-10). They are the inverted truth hierarchy. ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/HQ_20261105.md §5
10 Two rules that keep everything else honest: gates report, never kill; identity is the PDI md5, never reg 0x100. ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/MUST_PASS_VERIFICATION_LIST.md, ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/scripts/preflash_gate.sh

Answer format, any of these is fine:

DECISION: B — hold until asgard is back
DECISION: A, because the card is cheap to recover
DECISION: 1A 2C 3B 4A 5B          # all at once, by number

Where this file lives (absolute):

copy absolute path
canonical ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/Q_AND_A_20261105.md
beside the review that prompted it ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/reports/results/Q_AND_A_20261105.md
authoring copy ~/Desktop/research/eagle/Q_AND_A_20261105.md

The reply file it sits next to: ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/reports/results/STATUS_REVIEW_CORRECTED_20261105.md (box1 = [private-ip] per ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/build_machine_setup/README.md:593).

Cross-references: ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/HQ_20261105.md §7 · ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/OPERATOR_CONTEXT_20261105.md §11 · evidence in ~/Desktop/research/eagle/review_20261105/ (local) and ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/reports/results/ (repo).


Q1. May E2b (0xE4B713E2) be flashed onto talon2 17/3d?

What is proven. First timing-clean L512 LUT-GEMM ATT image: ^FINAL_TRUE WNS=+0.043 WHS=+0.012 (bar +0.030/+0.010), SLL SLR1↔SLR0 8,986 of 18,870 (47.6 %), PDI 2540c725ff966cfc89571375b98a71c3, 76,725,008 B, magic 16a5f792, same netlist as 0xE4B713E0 (reg100 0xE4B711E0). Preflash receipt: n_img_fail=0, n_img_unver=4, rc=3 BLOCKED_IMAGE_UNVERIFIED — nothing failed, four things could not be read.

What is not proven.

# unverified why it is stuck what closes it
1 the AMC firmware in the FPT is 0x13DE's (amc.elf cd6cc996, fpt.bin 45f47718), not E0's own asgard unreachable since 11-05 07:47Z asgard back, or a rebuild with E0's FW
2 E0's in-pipeline gate rows (GATE_VERDICTS_* on asgard) unread same same
3 no full-drafter replay at LANES 512, and RET_SER=1 has never been in one L512 replay lives on darehard, unreadable darehard/laptop back, or a talon2 replay after flash
4 host zero-bank load + G6 per-rank A8 scale ruling (same as 0x13DE) needs a decision, not a build your call, below

Why it might be moot within hours. 0xE4B713EB on buildx03 is the same netlist with E0's own floorplan and no BalancedSLR; it placed at WNS +0.124 / WHS −0.436 and is routing now. If it closes, it supersedes E2b and carries a second chance at a clean FW path.

option cost risk
A flash E2b on 17/3d at the next cold AC burns the slot + ~40 min borrowed FW on a JTAG-recovery-only path if it wedges
B hold E2b until asgard returns, then decide E2b vs EB ~0 now; you lose nothing that EB is not already testing nothing new learned about L512 tonight ◀ my recommendation
C rebuild E2b when asgard is up so it carries E0's own FW 1 build + the queue behind EB delays the only L512 image that closes
D other

DECISION: deferred — not mine to rule on until §0 of ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/LEARNING_BRIEF_20261105.md is corrected by the trantor agent.


Q2. Who may flash a quarantined-but-passing PDI?

The rules contradict each other, both written by you:

source says
run_arm.sh:40 (the runner, 2026-11-02) "flashable only after FINAL_TRUE ≥ +0.030/+0.010 + preflash_gate + USER decision"
CLAUDE.md (hard rules) PDI builds and -p 0 flash are pre-authorized once box1 has adjudicated

Every fan-out arm lands in that gap. Pick the rule that should win:

option effect
A ALWAYS explicit — every flash asks you safest, slowest; you become the queue
B pre-authorized iff receipt rc=0 PASS AND no .QUARANTINED marker keeps both rules: the gate decides, you only see exceptions ◀ my recommendation
C pre-authorized whenever n_img_fail=0, even if rc=3 fastest; would have flashed E2b tonight with 4 unknowns
D per-arm, declared in the slot's arm.env flexible, but the rule moves per build

DECISION: deferred — not mine to rule on until §0 of ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/LEARNING_BRIEF_20261105.md is corrected by the trantor agent.


Q3. One cold-AC slot, two candidates: which goes first?

candidate what it buys what it costs if it fails
E2b 0xE4B713E2 first L512 LUT-GEMM ATT on silicon at all the slot, plus a possible JTAG recovery
0x1342 3e82ea1d (staged since 11-04) the clock-arm image already adjudicated and waiting the slot, and E2b slips another day/AC
option
A E2b first, only after asgard returns and EB has a verdict ◀ my recommendation — one slot, and EB may hand you a better E2b
B 0x1342 first (it has waited longest, it is fully adjudicated)
C both in one session if the first boots clean

DECISION: deferred — not mine to rule on until §0 of ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/LEARNING_BRIEF_20261105.md is corrected by the trantor agent.


Q4. Which verilog branch is canonical?

fact value
.gitmodules points at tree-fsm-stage1
the checkout that built every shipped image shapert-4c1-20261021
uncommitted at last count 123 dirty files
unapplied stash 2026-08-26, 54 files
option
A declare shapert-4c1-20261021 canonical, freeze it as the shipping branch, mark the submodule pointer historical ◀ my recommendation — it is what the silicon actually came from
B rebase tree-fsm-stage1 onto it
C leave it, keep building from whatever is checked out

DECISION: deferred — not mine to rule on until §0 of ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/LEARNING_BRIEF_20261105.md is corrected by the trantor agent.


Q5. Publish the eleven-day hole on the research site?

The site has 141 posts and 54 diagrams, and stops at 11-04. Everything from 10-26 (7-card board) through tonight (RCA, fix B/C, L512 floorplan, E2b) exists only as repo markdown.

option effort
A 5 short posts, one per report (RCA · fix B/C · L512 wall+fix · shadow/JIT diagnosis · E2b closure) one push_page.py run each, ~40 min total ◀ my recommendation
B one long "10-26 → 11-05" campaign post ~20 min, less navigable
C regenerate the arch-book diagrams too needs the diagram generator, not in this repo

DECISION: deferred — not mine to rule on until §0 of ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/LEARNING_BRIEF_20261105.md is corrected by the trantor agent.


Q6. Is the Nov-4 rank-128 head idea a directive or a hypothesis?

CONCLUSIONS_20261104.md / fpga_lut_engine_20261104.md (this Mac) argue the drafter is ~18 % of round cost vs 82 % verify, and propose dropping the rank-128 head to save 3.3× bytes. That is an argument with numbers, not a measurement, and it changes the datapath.

option
A hypothesis — gate it behind a measurement before any RTL ◀ my recommendation
B directive — start the design now
C park it until the 7-card programme has a number

DECISION: deferred — not mine to rule on until §0 of ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/LEARNING_BRIEF_20261105.md is corrected by the trantor agent.


Answered log

date (UTC) # state unblocks when
2026-11-06 Q1–Q6 deferred the learning brief §0 is corrected, so the trade-offs are stated in the machine's real terms

The provisional calls written earlier (hold E2b, rc=0 pre-authorization, B before C, branch canonicalisation, five posts, rank-128 as hypothesis) are withdrawn from this file: they were made before the architecture and dev cycle could be stated cleanly. They stay in the git history at a509515f if the reasoning is wanted later.

Written 2026-11-05 20:40 EST. Facts trace to ~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/docs/HQ_20261105.md §4; the four E2b unknowns are itemised in the image's own *.QUARANTINED marker on box1 and in its *.PREFLASH_RECEIPT.

Q1 update — 2026-11-07 04:35Z (box1-qwen-junior; append-only)

The Q1 hold condition ("hold until asgard returns and EB has a verdict") is now met. Both D9 L512 images compared:

E2b 0xE4B713E2 EB 0xE4B713EB
Timing FINAL_TRUE +0.043 / +0.012 (bar +0.030 / +0.010) ROUTED −2.074 / 0.000, after 27.5 h of route; FINAL_TRUE still to come
PDI md5 2540c725ff966cfc89571375b98a71c3, 76,725,008 B none flashable
Open unknowns none: firmware (§9), RET_SER (§10) and L512 full-drafter replay (§11) closed in reports/results/QWEN27B_DRAFTER_20261020/lutgemm_d9_evidence/ADJUDICATION_0xE4B713E2.md —
preflash_gate rc=3: quarantine marker (needs the user), PDI-6 strap evidence and PDI-7 seam check, both also unverified on the silicon-proven 0x13DE —

The decision for the user: flash E2b to talon2 cards 17/3d (kit staged, run gated on D9E2_FLASH_GO=<md5>), or keep holding. EB is diagnostic only.