fpga-archive-reports-arch · EAGLE4 FPGA archive

Source reports/results/DRAFTER_SPEED_RESEARCH_20261107/FASTCLK_BUILD_0xE4B713EF_20261117.md · source md5 58af1f0fe82cf9002674280a04b03e04 · archived 2026-11-18

Download PDFOpen PDFRaw markdown

0xE4B713EF AFC8: ATT array fast-clock image, derived and LAUNCHED (box1-two-v2 build worker, 2026-11-17)

Program 1: one Llama-3.2-3B drafter on 7 V80s (ATT TP4 x4 + MLPV x2; H200 verifies). Motto: never deny the idea; engineer toward the fastest flashable PDI. Lever: ATT_FASTCLK_20261117.md (commit 6f16e0449). Base image: NFA8 0xE4B713CE (COMBO3_NF_A8QOVL_20261117.md, AUDIT_NFA8_0x13CE_20261117.md). Nothing was flashed, killed or removed. The NFA8 build (buildx03 s3LAtp4CE) and the A8OV build (build00 s3LAtp4D3) were not touched; I only read from them. NO FLASH before adjudication.

Calculations and hypothesis (everything that is not a measurement)

1. Image [M]

field value
build id 0xE4B713EF (label), probe 8'hEF, reg100 short id 0xE4B711EF
arm AFC8 = NFA8 0xE4B713CE + ATT_ARRAY_FAST_CLK_V 1'b1
host / slot buildx02 / s3LAtp4EF (clone /shared/build/s3LAtp4EF/aved_oneflip_p2s_head_asgard)
stage build host /shared/build/lutg_d9_tp4stk_EF_stage; box1 copy /mnt/data/eagle4_scratch/att_fastclk_build_20261117/lutg_d9_tp4stk_EF_stage
launch 2026-11-17T21:22:21Z: cd <repo> && TC3A_SLOT=s3LAtp4EF NFA8_MAXTHREADS=8 setsid nohup bash /shared/build/lutg_d9_tp4stk_EF_stage/launch_COMBO3.sh > .../afc8_pipeline.log 2>&1 < /dev/null &
first state STEP0 / STEP0.5 passed, including the CERT_TOP / CFG / GSAS / DCW checks that gate_fail on a mismatch. STEP1 PACKAGE was admitted at 21:22:22Z, and Vivado pid 3632834 was elaborating at 21:27Z with 0 ERROR lines.
pipeline logs (build host) <repo>/reports/results/UNI_DUALCLK_s3LAtp4EF_head_ASGARD_PIPELINE.log, GATE_VERDICTS_DUALCLK_s3LAtp4EF_head_ASGARD.txt

Host checks before launch (buildx02):

2. Build-id collision check [M]

The method reuses scripts/4c1/qgptq2/qg_id_check.py sources(): every non-QGPTQ2 c4_regmap int, the pre-Qwen c4_regmap, TRACKLIST, NOW.md, git log --all, every reports/docs *.md, tested_archive and the artifact file-name trees. I scanned each free low byte in the box1-two-v2 blocks 0x13B1-0x13FF that is not already in box1two_reserved_ids_20261116.txt.

3. Derivation [M]

Deriver: scripts/derive_lutg_tp4_nfa8_fastclk.py (24475912). Log: att_fastclk_20261117/build_EF/derive_EF.out (920089e6). Record: DERIVATION_AFC8.txt (d5c2ecc5).

  1. It runs derive_lutg_tp4_combo3nfa8.py (d042b8cd, NFA8_NF=1) unmodified with 0xE4B713EF / EF. All of its own PROOF-REPRO checks re-pass. Evidence: 772 / 772 files equal to the launched CE stage once the id, probe, slot and stage-name tokens are mapped (0 differ, 0 missing). The 6 NFA8 RTL files match.
  2. PROOF-REPRO: patch_att_fastclk.py (7ca95bf3) run on the staged NFA8 files, with --bridge = the fork's gemv_array_dualclk.sv 0395872d, reproduces all 5 / 5 fork files byte-exact. The same bridge md5 is present in tp4mlpv_134d, mlpv_1320, reimpl1376 and g8own_v2 MLPV stages. Then the fork files are overlaid, with the originals kept as *.pre_fastclk. Full tree: 149 / 149 non-backup files under fork src_full equal the staged copies.
  3. Strap: the staged config changes in exactly one line, ATT_ARRAY_FAST_CLK_V 1'b0 -> 1'b1. This is the F1_L100 sim arm's -d ATT_ARRAY_FAST_CLK_V=1. ARRAY_FAST_CLK_V 1'b0, ARRAY_FAST_TP_V 1'b0 and ATT_FASTCLK_RED_V 0 stay as residency. The top parameter ARRAY_FAST_CLK stays 0 (report section 5); the XCI row ARRAY_FAST_CLK want 0 is unchanged.
  4. I did not run a separate xsim of the staged copy. The staged RTL is byte-identical to the fork, and its one config difference is the define the F1 identity arm already used (13a16bea).
  5. must_pass: --stage PASS (A2: 0 macro-ternary port widths, 2 parameter-ternary) and --reddef PASS. A4 (pkg-param coverage) does not apply because the strap is a header define compiled from the packaged header; packaging visibility is checked by the XCI PARAM_VALUE.ATT_ARRAY_FAST_CLK row. test_record_profile_identity.py was not run because no host file changed.
staged file NFA8 (CE) AFC8 (EF)
top/eagle4_config.svh 89565763 38e7a756 (fork 6a58f888 + strap 1)
top/eagle4_drafter.sv 92ab14a2 901c191a
top/gemv_stage_array_scaled_slr.sv 7f754cc3 640a4530
primitives/lutg_a8q.sv 68be50f9 31d6dfdd
top/gemv_array_dualclk.sv a44f3e9f 0395872d (record MLPV 0x1361 bridge)
layers/premlp_normfuse.sv ddbc0795 ddbc0795
uni_pipeline_lutg_d9.sh 5a4b168b ebaac8da
run_lutg_d9.sh 65211bd9 b5de36be (CERT_TOP 901c191a / CFG 38e7a756 / GSAS 640a4530)
MD5LIST_head_hxf_cons.txt 6ecf582c bb89a4ee (392 lines; 392/392 OK on box1 and buildx02)
check_v1_netlist_lutg.sh 5ff9db77 b5352856
constraints_sllfix/eagle4_drafter_dualclk_loc_pblock.xdc fe2a53bd 7c1caf10
launch_COMBO3.sh / COMMON_MD5.txt 9d89ad8f / ad941439 (COMMON OK)
STAGE_SHA_0xE4B713EF.txt 4c69f53d

The full stage diff is in att_fastclk_20261117/build_EF/NFA8_STAGE_TO_AFC8_STAGE.diff (2cd39e2b).

4. Clock change [M]

5. Report-only rows added or re-keyed [R]

row wants
ATTAFC-staged (STEP3) staged ATT_ARRAY_FAST_CLK_V 1'b1 x1 and 1'b0 x0, ATT_FASTCLK_RED_V 0 x1, ARRAY_FAST_CLK_V 1'b0 x1, bridge 0395872d, drafter 901c191a
NFA8-staged lutg_a8q / array md5 re-pinned to 31d6dfdd / 640a4530
XCI PARAM_VALUE ATT_ARRAY_FAST_CLK 1, ATT_FASTCLK_RED 0
ATTAFC-SIBW-absent-bridge-present-netlist sibw_pump_e3 0, u_lane_s 0, gemv_dualclk_bridge >= 1, g_lcache >= 1
ATTAFC-netlist u_dc_bridge, u_a8q_s, g_lc_grp >= 1; g_a8q.u_lead and g_a8q.u_a8q 0
ATTAFC-fastclk-loaded routed clock_reset_clk_usr_1 endpoints >= 10,000 (replaces LUTG-pumpclk-unloaded, < 1000)
check_v1_netlist_lutg.sh SIBW-off gemv_dualclk_bridge > 0

Known inherited FAIL rows on this lineage are also printed on 0x13CE: p2s_selfref selftest SKIP, 8-3848, PREMLPCHECK, OSPREYCHECK, enforce-manifest, portw-reg0-range and THREEC_A STEP3. On 0x13CE they account for the 12 FAIL lines at the pkg / STEP3 stage (pkg-census contributes 4 of them).

6. Watcher [M]

7. Files

file md5
scripts/derive_lutg_tp4_nfa8_fastclk.py 24475912
scripts/att_fastclk_20261117/watch_afc8.sh, afc8_census.tcl 3cf06efc, 4e7ab64e
att_fastclk_20261117/build_EF/{STAGE_SHA_0xE4B713EF.txt, DERIVATION_AFC8.txt, derive_EF.out, NFA8_STAGE_TO_AFC8_STAGE.diff} 4c69f53d, d5c2ecc5, 920089e6, 2cd39e2b
scripts/4c1/box1two_reserved_ids_20261116.txt (+0xE4B713EF), 4C1_DELIVERY_TRACKER.md (+ launch line)

Housekeeping

8. Closure-recipe sibling 0xE4B7137F AFC8X (LAUNCHED buildx01 2026-11-17T21:46:05Z) [M]

Ordered by box1-two-v2 to cut closure risk on EF. The OOC margin was +0.118 / +0.010.

Stage

Recipe diff vs EF (att_fastclk_20261117/build_7F/EF_TO_7F_RECIPE.diff, 5abec301)

knob EF 0x13EF AFC8X 0x137F why
place_design -subdirective {Floorplan.ForceSpreading.high GPlace.ForceSpreading.high GPlace.ReduceCongestion.high DPlace.ReducePinDensity.high} -directive Explore + the same subdirectives Lineage evidence: BC-ETO re-placed 0x13BC's own opt dcp with exactly this delta and reached FINAL_TRUE +0.050 / +0.011, vs the 0x13BC pipeline's +0.040 / +0.010 (G6_REISSUE_BC_ADJUDICATION_20261111.md:40-42). The 0x13BC loss was congestion, so the spreading set stays. ExtraTimingOpt is not a Versal 2025.1 place directive (4-2308 revert). Project mode: DIRECTIVE Explore and SUBDIRECTIVE are accepted together (buildx01 /shared/scratch/afc8x_dirtest_20261117/t.log, rc 0).
STEP3 "place DIRECTIVE set" check gate_fail report-only row AFC8X-place-directive (wants DIRECTIVE Explore x1)
in-build setup margin, clock_reset_clk_usr_1 (132 MHz) 0.040 0.080 The fast clock is the thin domain. cancel_hook_d1.tcl (unchanged) zeroes it before the final report and write_device_image, so FINAL_TRUE is unmargined. All other clocks stay at 0.040 / 0.020. Margin XDC 3c5ebe84, md5 re-pinned.
route / phys_opt / threads / floorplan AggressiveExplore / AggressiveExplore / 8 / sllfix XDC unchanged

Pipeline file md5s: pipeline 3eb67f1b, run ffe6be54.

Host buildx01 at launch

Build id

Watcher

9. Verdict on the two EF-only FAIL rows (2026-11-18 ~00:35Z): both are check misfires, not defects in the image [M]

0x13EF has 13 FAIL rows in GATE_VERDICTS, and so does 0x137F. 11 of them are inherited and appear identically on 0x13CE. The two below are new on EF / 7F, and both are mis-keyed checks. Neither build was touched, and no fixed sibling is needed for these two rows.

(1) ATTAFC-staged: drafter b82eaad0, but the row wants 901c191a. This is my row, and it was mis-authored.

(2) STEP3: staged gemv_stage_array_scaled_slr.sv lacks the g_lutg / g_a8q arms. This is an inherited text check that the fast-clock patch is designed to break.

Why both builds took one STEP4b retry. The BD reported ap_clk_hbm 285,999,744 against the seed 285,999,740. ap_clk 65,999,940 and the fast clock 131,999,881 matched their seeds.

Deciding evidence still to come: the afc8_census.tcl rows on the opt dcp.

Fix for future siblings (not applied to the running builds):