fpga-archive-demo-box1-20261110 · EAGLE4 FPGA archive
Box1 4-card demo: temperatures under a humaneval benchmark (2026-11-10)
Result: the humaneval run did NOT happen. The cards are brought up and armed. The golden level identity vs the E2b record passes. But the SGLang benchmark refused at the first tree, before any request was served. Its kit requires a bf16 golden.json PASS, and the E2b image (LUTG W2×A8) never passes that gate. Only idle temperatures and an 85 s load-free window exist. On the coordinator's relay of the user ruling, box1 card work moved to box1-qwen-junior, so no re-run was attempted.
What is on the box1 cards
Identity evidence:
/var/log/auth.log, 2026-11-07T20:18:48-05:00:ami_tool cfgmem_program -p 0×4.- 01/42 ←
oneflip_p2s_head_fan_0xE4B713E2_66m_2540c725.pdi - 61/62 ←
oneflip_p2s_tail_asgard_0xE4B71342_66m_3e82ea1d.pdi
- 01/42 ←
~/flash_p0_20261107/SUMMARY.txt: rc=0 ×4 at 01:54–01:56Z on 11-08; activated by the 11-08 cold AC.- Readback on 11-10:
- 01/42: reg100
0xe4b711e0, tp_bid0xe4b713e0; - 61/62: reg100
0xe4b71142, tp_bid0xe4b71342.
- 01/42: reg100
Also in the same auth.log: a -p 1 cfgmem_program on 61:00.0 at 2026-11-07T19:44:37, pkill'd at 19:54:50, before the -p 0 reflash. The flashing session is unknown; box1-qwen-junior says it was not them.
Bring-up (run dir /mnt/data/eagle4_scratch/4c1_runs/box1_e2b_tempbench_20261110)
| step | result |
|---|---|
| discover | OK. Needed the new flash record scripts/4c1/flash_records/flash_4c1_box1_e2b_20261107.json and the 0x1342 MLPV row in c4_regmap.py (commit 9d56c6c0c). |
| weight verify | 0x13E2 Llama: box1 log lutgemm_d9e2_20261106/logs/verify_att_lutg_13E2_llama_20261106.log, 2/2 PASS + RED. 0x1342: re-verified on box1 (box1_temp_20261109/verify_mlpv_1342_box1_20261110.log), 2/2 PASS + RED. The box1 boot MANIFESTs are md5-identical to talon2's 4c1_boot_llama (ef71d0a2 / 8af9f2f9 / 1c8e12d7 / 8378ee10). |
| linkfix / link | NOTHING_FROZEN; LINK OK (all ports up) |
| boot | OK, after porting talon2's rule that accepts 0x80 = 0xF5000200 as fresh (commit 5337e68b7) |
| arm | OK, depth 15 |
| golden (bf16, he4d15cr, md 15,7,2) | FAIL, 0/12 (L0 8/8, divergence from L1) — expected for LUTG A8 |
golden level identity vs E2b record TRACK1_RECORD_F3T_20261109/e2b_same_day/e2b_record_20261106/golden_llama_e2.json (md5 399b0e06), via cmp_golden_levels.py |
trees identical 12/12, levels 56, rc 0 = the same gate F3t passed |
Benchmark attempt
Command:
RUN=<run dir> BENCH=.../humaneval/question.jsonl MAXREQ=80 MAXNEW=256 DEPTH=7 WIDTH=2 TAG=4c1_box1_e2b_he80x256_temps bash scripts/4c1/run_4c1_sglang.sh
- Window: 00:20:11Z → 00:21:36Z.
- It died at the first prefill:
RemoteCardError: ... golden.json verdict 'FAIL' ... run the golden gate in THIS run first (EAGLE4_4C1_ALLOW_UNGATED ...). - Log:
reports/results/FOURCARD_DESIGN_20261016/4c1_runs/4c1_box1_e2b_he80x256_temps_20261109_192011.log. - tok/s and accept: none (no request served).
Temperatures
Source: /mnt/data/eagle4_scratch/box1_temp_20261109/temps.csv (5 s, every Alveo hwmon sensor + H200). The logger is still running (6 h from 00:04Z). Analysis: analyze.py temps.csv <t0> <t1>.
| card | idle max °C | max °C in the 85 s window (cards idle) | per-sensor (window) |
|---|---|---|---|
| 01 (ATT0) | 53 | 54 | 38 36 42 46 46 0 53 54 52 35 45 |
| 42 (ATT1) | 53 | 53 | 37 37 42 46 46 0 53 53 52 35 45 |
| 61 (MLPV1) | 57 | 57 | 40 38 46 50 49 0 57 57 56 37 49 |
| 62 (MLPV0) | 62 | 62 | 43 40 50 54 53 0 61 62 59 39 53 |
H200: idle 51 °C / 73 W. During model load + capture: peak 65 °C, 517 W.
Open for the box1 owner (box1-qwen-junior)
To get loaded temps, re-run the command above with the gate handled the way talon2's E2b serving runs handle it. Either:
- write a golden.json PASS from the record level-identity, or
- set
EAGLE4_4C1_ALLOW_UNGATED, which talon2's PQ+C kit may set.
The cards are still armed at d15, and the temp logger is still writing.