fpga-archive-reports-record · EAGLE4 FPGA archive
0xE4B713D6 (COMBO3) + 0xE4B713DB (D5F): flash adjudication of the builds' OWN pipeline PDIs, 2026-11-16
Program 1: the 7-card Llama-3.2-3B drafter on talon2 (ATT TP4 ranks 17/3d/bd/e1). Motto: engineer toward the fastest flashable PDI. Adjudication worker for box1-two-v2. Read-only on both builds. Every Vivado run used a create-only COPY of the dcps (src md5 == dst md5). Nothing was flashed, killed, moved or removed. The rescues M13D6A (buildzd04) and M13DBA (buildzd01) were not touched. Tags: [M] measured (the artifact is named), [R] read.
Verdict: both GO (-p 0, the same md5 on all four ATT ranks). Each image is staged on talon2. Neither is flashed.
| COMBO3 0xE4B713D6 | D5F 0xE4B713DB | bar | |
|---|---|---|---|
| levers | 0x13BC + G5_P2W 8 + R3_EARLY 1 + KV1H2C 1 + HANDOFF_CUT 1 (no KVRD credit fix) | 0x13D5 + KVRD_CRED_FIX_V 1 | |
final FPT PDI (build root amd_v80_gen5x8_25.1.pdi == ~/eagle4_artifacts/uni_oneflip/ copy) |
d2e737d7c7134cb0d317345a6e405b1b, 79,827,088 B |
24849a8f27620211a545143653871ddc, 81,617,760 B |
never impl_1/top_wrapper.pdi |
| final dcp (the in-memory state write_device_image used; post-route phys_opt "netlist was not modified" on both) | impl_1/top_wrapper_postroute_physopt.dcp 7651f0988cebc9af96d713fb08755c02 |
same name bfbf3776ac69f72c76ca62f3891f1c10 | |
| opt dcp (srpol tooth 2) | top_wrapper_opt.dcp 23bbb321e9f6b7d936d8b494b1463836 |
e007241aeb033d87ae99abf43ca9ac78 | |
| FINAL_TRUE (margin 0) | +0.044 / +0.010, TNS 0 / THS 0, 0 / 0 failing of 2,185,430 / 2,184,339 endpoints | +0.042 / +0.011, 0 / 0, 0 / 0 failing of 2,182,926 / 2,181,835 | >= +0.030 / >= +0.010 |
| as-saved (fresh open) / margined (+0.040 setup / +0.020 usr hold re-applied) | +0.044 / +0.010; +0.004 / +0.010 | +0.042 / +0.011; +0.002 / +0.010 | |
| worst setup | ch0_txoutclk_5 Aurora aur_link3/gtm_tx_gearbox count -> data_in_d1 (4 levels, req 2.482) |
clk_pl_2 CPM QDMA AXIST.u_demux ch_crdt -> c2h in_data CE (3 levels, req 4.000) |
|
| worst hold | clkout1_primitive tp_x/u_hc/u_s1/u_in async-fifo doutb pipe |
ch0_rxoutclk_2 aur_link descrambler -> cbcc lane3 |
|
| route (fresh report_route_status) | 2,208,786 / 2,208,786 fully routed, 0 errors | 2,220,959 / 2,220,959, 0 errors | 0 |
| DRC (fresh report_drc) | 20,833 checks, all Warning; 0 Error / 0 Critical | 20,833 checks, all Warning; 0 Error / 0 Critical | 0 errors |
| magic / AMC | 16a5f792; AMC_IN_BAND_DBG PDI 32 == build/amc.elf 32 (amc.elf e39ac04917bb2988ece765249ab35b11) |
16a5f792; 32 == 32 (amc.elf 9ab4329e64d7db96c3dd8e85d9a7876b) |
equal, != 0 |
srpol --srpol (census tcl d56fe71b) |
PASS 0 mixed, selftest PASS, FDRE 356,014 inv0 + 568,694 inv1 = 924,708 censused (buildzd00) | PASS 0 mixed, selftest PASS (buildxd05) | 0 mixed |
srpol tooth 2 --srpol-optdiff (fast 050586f4/f1f87afe, resolver 0a0bd072) |
PASS changed 0, rep_diff 0, root_changed 0; common 932,157; candidates 10, all reset_absorbed (the 10 u_pipe_bridge/rd_chunk_idx_reg[0/1]_replica_*: R tied off and the same usr_0_psr root drives the D-LUT, a control-set reduction the tool rules legit) (buildxd03) |
PASS changed 0, rep_diff 0, root_changed 0; common 931,953; candidates 0 (buildxd04) | changed 0 |
KV-reader NoC (ncr_scan.py 27f3f4da on impl_1/top_wrapper.ncr) |
ncr 0c8f8a6c: K NMU512_X0Y3 / V NMU512_X1Y7, shared_READ_out_links 0 = byte-for-byte the BC / BE placement (both 0 aborts on silicon) | ncr 608ff208: K X0Y4 / V X1Y7, 0 shared (and the credit fix is in the netlist regardless) | COMBO3: 0 |
Hold: COMBO3's +0.010 meets the bar exactly, as BC 0x13BC did (+0.040 / +0.010, GO 11-11). That +0.010 is on tp_x link-shim FIFO logic (clkout1_primitive), the class of hold BC / BE / BA / D5 all closed at +0.010-0.011. Drafter clocks are not the limiter on either image.
Margin arithmetic: both builds carried KVC2 margin_att_d1.xdc (setup 0.040 on 13 clocks, hold 0.020 on usr_0..2). The census numbers +0.004 / +0.002 were read from post_route_phys_opt_report_timing_summary_0.rpt, which is written BEFORE cancel_hook_d1.tcl (D6 runme.log:3542-3544, DB :2549-2551). write_device_image ran after the cancel (write_device_image.pre.tcl prints ^KVC2_MARGIN_CANCELLED clocks=2 again). ^UNC_ASSAVED is empty on all 12 listed clocks of each fresh-opened copy. This confirms the rescue agent's p0_src numbers.
Per-clock FINAL_TRUE [M] (^SETUP_BY_CLK / ^HOLD_BY_CLK, ft_stdout.txt)
COMBO3: clk_pl_0 +0.286 / +0.013 · clk_pl_1 +4.229 / +0.069 · clk_pl_2 +0.045 / +0.011 · clkout1_primitive +0.052 / +0.010 · global setup worst = ch0_txoutclk_5 +0.044.
D5F: clk_pl_0 +2.720 / +0.018 · clk_pl_1 +4.083 / +0.055 · clk_pl_2 +0.042 / +0.023 · global hold worst = ch0_rxoutclk_2 +0.011.
The full lists (with the drafter usr clocks and the GT clocks) are in each host's ft_stdout.txt, ft_<id>_setup20.txt and ft_<id>_hold20.txt.
GATE_VERDICTS / marker rows [M]
- Both markers: NFAIL 39 (parent-shell 27). Sidecars: COMBO3
~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_asgard_0xE4B713D6_66m_d2e737d7.GATE_VERDICTS.txtmd5 441fd757 (331 lines). D5F..._0xE4B713DB_66m_24849a8f.GATE_VERDICTS.txtmd5 edaadc96 (327 lines). - Normalised FAIL rows (timestamps / slot / id / tmp dirs stripped,
sort -u): COMBO3 40 == D5F 40, apart from the pblockgate count (22 vs 23). Against the record 0x13BE (38 rows), the diff is 4 rows on each image, the same 4 rows ruled benign for 0x13D5 (ADJUDICATION_0x13D5_20261113.mdsection 3):8-3848-bake-drop (1)=Net score_buf in attention_head_assembly_pipe does not have driver(logs/dualclk_head_asgard_pkg_synth.log:1635,attention_head_assembly_pipe.sv:316). This is by construction of G5_P2W > 1 and fires on every P2W image.strap-watcherFAIL. It is a race with the sim-netlist write: COMBO3 FAIL 01:16:59Z, netlist mtime 01:17:54Z, thenstrapgate VERDICT: PASS+strap watcher: PASS01:18:52Z (DEPTHN_GATE_WATCH.log:4864,4895,4951). D5F FAIL 05:40:16Z, netlist 05:41:08Z, PASS 05:41:15Z (:3094,3125,3181). The finished netlists holdg_depth_n864 / 872 andemb_gather866 / 874.tree_level_fsmis 0, the ATT role prune, as on BC.- pblockgate dead filters 22 / 23 (vs 21): modules the ATT role does not have.
SIBW-TP tp_ready_r=32(vs 37): informational, the B4 count.
- Every other row is shared with the GO'd record image (OSPREY ATTN_N_HEADS 8-vs-32, CP1Q link gates, enforce-manifest vs 0x1128, etc.: the ATT-family standing set). 0 flash-relevant rows.
- Lever residency PASS rows. COMBO3:
lutg-keep-G5_P2W_V 8,KV1H2C_V 1,R3_EARLY_V 1,KV1H2C-netlist voff_r_fd=31 kvf_voff_q_fd=31 kvf_auto_q_fd=1,hcut-link HANDOFF_CUT=1,hcut-config,R3EARLY-wrapper .R3_EARLY(1),G5P2W-wrapper .G5_P2W(8). D5F:G5_P2W_V 8,R3_EARLY_V 1,KVRD-CREDFIX-staged(reader 4aa93eb0, cap x1,.CRED_FIX(KVF_CRED_FIX)x1), plus both wrapper rows. - The build's own route reports agree:
route_report_route_status_0.rpt0 routing errors on both.
Silicon notes for the window (not part of the static bar)
- COMBO3 carries D5's G5_P2W 8 + R3_EARLY 1, and D5 aborted its KV reader on 2/2 boots (0x1D4 0x01013000). D5's ncr had 18 shared K/V READ links (BA 21). COMBO3's placement is the 0-shared BC / BE one, so the RCA's trigger class is absent. COMBO3 has no credit fix, though: an abort on COMBO3 would refute the link-sharing explanation. Run serve-first at d15 with 0x1D4 / 0x1E8 read after every serve on all 4 ranks, exactly as
ADJUDICATION_0x13D5_20261113.mdsection 6 (a)-(f). Restore 0x13BE a21ca534 on any abort. - D5F lands on the clean class too (0 shared), so silicon will not isolate the credit fix. Its value is as the D5-lever image that carries the fix.
- Host ABI: COMBO3 = 0x13C3 (CSR 0x228 KVF_VOFF) + 0x13BE (link-shim BELL live regions). D5F = 0x13D5 (no CSR change). Both kits need their IMAGES rows / boot roots before arm. Neither row exists yet: only EXPECT_PDI_MD5 is added.
- Golden: run
golden_fp_gate.py(ref 51524c4f). "rc 50 / gate 0" alone is vacuous on 7-card.
Staged / recorded
- talon2
~/eagle4_talon2_4c1/att_combo_20261116/(create-only; streamed via box1, nothing stored on box1):oneflip_p2s_head_asgard_0xE4B713D6_66m_d2e737d7.pdimd5 d2e737d7... (79,827,088 B, magic 16a5f792, AMC 32) andoneflip_p2s_head_asgard_0xE4B713DB_66m_24849a8f.pdimd5 24849a8f... (81,617,760 B, 16a5f792, AMC 32). md5 re-verified on talon2. scripts/4c1/c4_regmap.pyEXPECT_PDI_MD5: +0xE4B713D6 d2e737d7..., +0xE4B713DB 24849a8f... (import check resolves both).
Evidence (all create-only, on the VMs)
| what | where | md5 |
|---|---|---|
| FINAL_TRUE engine | final_true_adjud.tcl (tp4_adjud.tcl FINAL_TRUE part + nfail + route_status + drc) |
f06c801c36f7e02190df102f7ecd4c3a |
| COMBO3 timing | buildzd03 /shared/scratch/adjud_13D6_20261116/ ft_stdout.txt / ft_13D6_tsum_final_true.rpt / _route_status.rpt / _drc.rpt |
8f13b2ac / e6eb8668 / 6b5cec66 / 9b46d92b |
| D5F timing | buildzd05 /shared/scratch/adjud_13DB_20261116/ same names |
ba8212a0 / 96f57618 / faaf55a6 / 8b112018 |
| srpol | buildzd00 / buildxd05 .../gate/adj<id>_srpol/adj<id>_srpol_srpol.txt + GATE_VERDICTS_adj<id>.txt |
7884321a / d41e6a71 |
| optdiff | buildxd03 / buildxd04 .../gate/adj<id>_optdiff/optdiff.txt (+ verdict appended to that host's GATE_VERDICTS_adj<id>.txt) |
38db75ec / bb6c1834 |
| gate tools | must_pass_gate.sh 475ed2dc (repo copy under /shared/scratch/adjud_mp_20261116/ on the 4 gate hosts, run locally with XILINXD_LICENSE_FILE=[[withheld]]; the VMs' default 2101/2100 hung a licence smoke at create_ip) |
Licence: every run used [[withheld]] (box1 relay of midyard's fixed lmgrd). No licence error in any log.