fpga-archive-reports-record · EAGLE4 FPGA archive
MLPV_G8_OWN full PDI builds: 0xE4B71361 + timing hedge 0xE4B71362 (2026-11-11)
Program 1: one Llama-3.2-3B drafter on 7 V80s (4 ATT TP4 + 2 MLPV on talon2). Motto: engineer toward the fastest flashable PDI, accuracy exact.
RTL: fork verilog/forks/mlpv_g8_own_20261111 v3 (MLPV_G8_OWN_RTL_20261111.md, audit AUDIT_MLPV_G8_OWN_20261111.md final verdict GO to build on v3).
Nothing flashed. Gates are report-only. Nothing removed.
Launch facts (0x1361; 0x1362 below)
| item | value |
|---|---|
| build id / probe | 0xE4B71361 / PROBE_REV 8'h61 (0x1361..0x1363: 0 hits in git log --all of repo + verilog, TRACKLIST, NOW.md, c4_regmap.py, /mnt/models stage dirs, and /shared/build on buildxd00-05 + buildx00-04, checked 22:2xZ) |
| host / slot | buildxd00 / s3LCg8own61 (own heavy lock /shared/build/_hostwide/oneflip_logs/slot_s3LCg8own61_own) |
| launched | 2026-11-11T22:38:12Z, PGID 2401244 (pipeline pid 2401245); MemAvailable 120.7 GB, no Vivado, load 0.01 right before launch |
| stage | local /mnt/data/eagle4_scratch/mlpv_g8own_20261111/4c1tp4_1361_g8own_mlpv_stage, remote buildxd00:/shared/build/4c1tp4_1361_g8own_mlpv_stage (remote MD5LIST + COMMON_MD5 md5sum -c OK) |
| launch line | LAUNCH_0xE4B71361.txt line 1 = the 0x134D launch line with stage path + slot swapped (run_4c1v12_mlpv_dg30.sh, then post_build_closure.sh) |
| provenance | /mnt/models/stage_E4B71361/STAGE_SHA_E4B71361.txt (copy in the stage dir) |
| derive tool | scripts/4c1/hedge/derive_1361_g8own.py (count-asserted, create-only; log STAGE_G8OWN_1361.log in the stage) |
Source
- Base = the 0x134D stage copy
/mnt/data/eagle4_scratch/c1_itempar_20261110/base_134d2. Measured byte-identical tobuildxd02:/shared/build/4c1tp4_134d2_mlpv_stage(620 files; the remote has 3 extra logs only). patch_g8_own.py(6f5fa204) applied to the STAGED copies, thenmlpv_g8_own_link.svcopied in, thenMLPV_G8_OWN_V 0 -> 1baked:
| file | 0x134D staged | 0x1361 staged | fork v3 |
|---|---|---|---|
| top/eagle4_drafter.sv | 17fa81ad | 845ca21c | 845ca21c |
| g8/g8_spec_frontend.sv | 914dcea4 | c359b44c | c359b44c |
| g8/g8_rescore_p.sv | 33ff4850 | 2cb84fc7 | 2cb84fc7 |
| layers/mlpv_vs_link.sv | 0cdea4f8 | 15dd933f | 15dd933f |
| layers/mlpv_g8_own_link.sv | (none) | ffd8d5b9 | ffd8d5b9 |
| top/eagle4_config.svh | 25668c69 | c7c25ed2 (baked) | d3c6ca27 (unbaked default 0) |
- The config hunk (staged 25668c69 -> patched) equals the fork hunk (parent 0ae1df37 -> v3 d3c6ca27): 5 +/- lines, asserted.
- Straps read back once each in the staged config:
MLPV_G8_OWN_V 1,S1_OVERLAP_V 0,MLPV_VSHARD_V 1,MLPV_XCH_V 1,FG8_V 1,G8_NCH_V 4. G8_SPLICE is 1 on the tail (STEP0.6a). So interlockg_g8_own_xis satisfied. - Pipeline edits (count-asserted): stage path x6;
PROBE_NEW 8'h61 / BUILD_ID_P 0xE4B71361; own lock; CERT_TOP_OSPREY -> 845ca21c; CERT_CFG_OSPREY -> c7c25ed2.mlpv_g8_own_link.svadded to the STEP0.6 add_files injection and to the STEP3 staged-IP file list (a missing file would black-box u_g8own at BD re-synth).G8OWN=1at bothcheck_v12_netlist.shcall sites, and a report-only G8OWN block incheck_v12_netlist.sh:- packaged config
MLPV_G8_OWN_V 1and 0 OWN_RED defines; - BD synth wrapper literal
MLPV_G8_OWN= 1; - netlist
g_g8own.u_g8ownpresent,g_g8own_off/g_g8own_tieabsent, 512 distinctf_mine_reg,mir_regpresent; - rx_full count as INFO only.
- packaged config
Pre-launch checks (all on the staged copies)
check_stage_g8own.sh <stage src_full> 1: STAGE_G8OWN PASS (10/10 rows: 6 RED macros undefined, H1 line, EXACT_END mask, own_link present, MLPV_G8_OWN_V = 1).- Audit S1 rows that the tooth lacks:
- staged drafter
.EXACT_END(MLPV_G8_OWN)x1,g_g8ownx5,.G8_OWN(MLPV_G8_OWN)x1,g_g8_own_xx1; - staged rescore
.G8_OWN(G8_OWN)x1; - all 4 staged RTL md5s equal the fork v3 md5s;
OWN_RED_in the stage tcl/sh/py/xdc/txt: 0 (the only hit is my own check row); the pkg tcl runsset_property verilog_define {}.
- staged drafter
gen_pkg_param_list.shon the staged config, run on both box1 and buildxd00: emitsMLPV_G8_OWN 1(146 parameters).- must_pass_gate:
--stage(A2) PASS;--reddef(A5) PASS (566 files). - Stale pre-staged copy
forks/tail_fgi_20260930/build/src_staged_oneflip_p2s_tail(drafter f87f0f9a, pre-G8_OWN) is in the stage. STEP0.5 replaces it (rsync -a --delete src_full/ STAGESRC/, pipeline :855). The in-build row below confirms the replacement.
Build A (0x1360 + G8_OWN): NOT BUILT, interlock
- 0x1360 staged config
1b372c2ebakesS1_OVERLAP_V 1(withBATCHMODE_V 1):STAGE_SHA_E4B71360.txt, config line 314. g_g8_own_x(fork drafter :1621) requires!S1_OVERLAP. Patching the 0x1360 drafter b2a825d4 anchors cleanly (patch_g8_own.py --checkOK), but the image would elaborateERROR_MLPV_G8_OWN_REQUIRES_..._S1_OVERLAP_0.- The interlock is real, not cosmetic. Under S1 overlap, child j+1's G9 runs while child j's score phase holds the TP link, so
mlpv_g9_xch_tp(2 tile slots) overflows (CHANGES.md "Composition with D7").- 0x1360 relaxes the other S1 interlocks with
|| BATCHMODE; G8_OWN has no proof for that mode.
- 0x1360 relaxes the other S1 interlocks with
- Not forced. A 0x1360-lineage G8_OWN image needs one of CHANGES.md fixes 1-4 first, or proof that BATCHMODE never lets G9 j+1 overlap score j.
- Consequence for the A/B: 0x1361 is 0x134D + G8_OWN, not 0x1360 + G8_OWN.
- 0x134D differs from the flashed 0x1360 in TREE_P (32 vs 64), S1_OVERLAP/BATCHMODE and TREEFIX.
- The clean silicon A/B is therefore 0x1361 vs 0x134D on the MLPV cards, not vs 0x1360.
- The second VM (build00) runs a real implementation variant of the same source, not a deterministic copy: 0x1362 below (coordinator request 22:4xZ).
0xE4B71362: timing hedge (same RTL, different place recipe)
| item | value |
|---|---|
| build id / probe | 0xE4B71362 / 8'h62 (free: same search as 0x1361) |
| host / slot | build00 / s3LCg8own62 (own lock slot_s3LCg8own62_own) |
| launched | 2026-11-11T22:49:35Z, PGID 2223283 (pipeline pid 2223284); MemAvailable 118.5 GB, 0 Vivado right before launch |
| stage | local /mnt/data/eagle4_scratch/mlpv_g8own_20261111/4c1tp4_1362_g8own_mlpv_stage, remote build00:/shared/build/4c1tp4_1362_g8own_mlpv_stage (remote md5 lists OK) |
| provenance | /mnt/models/stage_E4B71362/STAGE_SHA_E4B71362.txt |
| derive | scripts/4c1/hedge/derive_1362_g8own_place.py on the 0x1361 local stage |
- RTL bytes = 0x1361's. MD5LIST and COMMON_MD5 are unchanged; only PROBE_REV differs (STEP0.5 staged drafter 35327c33 vs 0x1361's 9f4998cc).
- The only implementation change:
STEPS.PLACE_DESIGN.ARGS.DIRECTIVE Exploreis added in the STEP3 impl injection.- It goes on top of the same ForceSpreading/ReduceCongestion/PinDensity SUBDIRECTIVE. That combo was verified rc 0 on Versal 2025.1 by the BC-ETO run (commit 38b02608f).
- The place-spread stage/impl gates still bind.
- phys_opt, post-route phys_opt and route stay AggressiveExplore.
- Why not ExtraNetDelay_high:
derive_mlpv_placearm.pylists Explore / AggressiveExplore / Default / RuntimeOptimized / Quick as the 2025.1 Versal-legal set. Explore is the one already proven in combination with the subdirectives. - Pipeline diff vs 0x1361 (pipeline md5 f10d8eee vs 6bbfada6), with id/slot/path normalised:
- probe;
- the one recipe line;
- two report-only rows,
place-arm-Explore-stage(build_design.tcl) andplace-arm-Explore-impl(impl_1/top_wrapper.tclplace_design -directive Explore). - The injection was tested on a dummy tcl.
- Same pre-launch checks: STAGE_G8OWN PASS; must_pass
--stagePASS and--reddefPASS. - The derive tool's last leftover assert tripped on a benign add_files comment (
# MLPV_G8_OWN 0xE4B71361) after every edit had landed.- I re-ran the leftover counts by hand: 0/0/0/0.
- The tool now exempts that comment. Re-running it REFUSEs because DST exists (create-only).
- c4_regmap.py: 0x1361 and 0x1362 added as PENDING:
- EXPECT_PDI_MD5 comment rows;
- MLPVTP4 card entries, reg100 0xE4B71161 / 0xE4B71162,
g8_own: True; - TP4_STAGED config c7c25ed2 / RINGB 1 / OFF 0x140000 / SOLO 1.
- The import works.
- Flash rule: only the build that adjudicates GO gets flashed (positive true timing on ALL clocks + AMC count == amc.elf + the quarantine rows).
Phase log (updated by artifact)
| UTC | phase | evidence |
|---|---|---|
| 22:38:12 | 0x1361 launched; fork md5 list OK, common md5 OK | 4c1v12_mlpv_pipeline.log lines 1-2 |
| 22:38:14 | 0x1361 STEP0.5 staged src OK: PROBE_REV 8'h61, 2-file delta (PROBE_REV + role config), staged drafter 9f4998cc (= 845ca21c + PROBE_REV line); staged own_link ffd8d5b9 + frontend c359b44c; staged config still MLPV_G8_OWN_V 1 x1; derived pkg tcl :479 adds mlpv_g8_own_link.sv; STEP1 PACKAGE start |
UNI_CONSOL_TAIL_tail_ASGARD_s3LCg8own61_PIPELINE.log (VM repo reports/results) |
| 22:38:13 | gate-selftest p2s_selfref_grep_selftest INCONCLUSIVE [FAIL] row: inherited, the same row is on 0x134D (buildxd02 GATE_VERDICTS 2026-11-09 16:56); not G8_OWN | GATE_VERDICTS |
| 22:49:36 | 0x1362 STEP0.5 staged src OK: PROBE_REV 8'h62, 2-file delta, staged drafter 35327c33; STEP1 PACKAGE start (same selftest row) | UNI_CONSOL_TAIL_tail_ASGARD_s3LCg8own62_PIPELINE.log on build00 |
| 00:25:48 | 0x1361 STEP1 packaging done (1 h 47 m): component.xml MODELPARAM_VALUE.MLPV_G8_OWN = 1 and PARAM_VALUE.MLPV_G8_OWN = 1; packaged src/ carries mlpv_g8_own_link.sv, and packaged config bakes MLPV_G8_OWN_V 1 x1. must_pass --pkg --kind MLPV (dual 0 and 1): PASS; A1 awaddr/araddr [8:0], A3 reaches 0x1FC, addressBlock 64 KiB |
buildxd00 packaged_ip_oneflip_p2s_tail/eagle4_drafter_depthn_1_0/component.xml (301,452 B) |
| 00:26:44 | 0x1361 STEP3 done (all STEP3 source/staged-IP rows OK incl. the own_link file list); STEP4 BUILD_ALL start (224 GB disk, heavy_impl admitted with 115 GB available) | pipeline log |
| 00:26:44 | 0x1361 GATE_VERDICTS FAIL rows so far = ENGCHECK, RESIDENCYCHECK, KVCLRCHECK (x2 pkg-census), tpsafe-loop-stage, gate-selftest. Every one is also on the parent 0x134D (g8own_1361/GATE_VERDICTS_PARENT_0x134D_s3LCtp4d2.txt, 185 lines, copied from buildxd02), so none is G8_OWN-caused. The adjudication diffs the final FAIL set against that parent set |
GATE_VERDICTS |
| 00:32:12 | 0x1362 STEP1 done: component.xml MLPV_G8_OWN = 1 (MODELPARAM + PARAM); own_link in src; must_pass --pkg PASS. It differs from 0x1361's only in 2 checksum lines (PROBE_REV) and the lmdown .mem glob order |
build00 component.xml |
| 00:33:24 | 0x1362 STEP3 done; place-arm-Explore-stage PASS (build_design.tcl has STEPS.PLACE_DESIGN.ARGS.DIRECTIVE Explore next to the SUBDIRECTIVE line); STEP4 BUILD_ALL start; FAIL-row count 9 = same set as 0x1361 |
build00 pipeline log + GATE_VERDICTS |
| 00:35:41 | 0x1361 BD synth wrapper top_eagle4_drafter_lb_0_0.sv: .MLPV_G8_OWN(1) (:573; also in the CONFIG string :56), .S1_OVERLAP(0), .FG8(1), .G8_SPLICE(1'D1), .G8_NCH(4). MLPV_VSHARD / MLPV_XCH are not wrapper ports; they come from the baked config defines (1/1), and a 0 would elaborate the g_g8_own_x error module |
buildxd00 prj.gen wrapper |
| 00:41:38 | 0x1362 BD synth wrapper: .MLPV_G8_OWN(1), .S1_OVERLAP(0) |
build00 prj.gen wrapper |
| 02:36:50 | 0x1361 drafter OOC synth done (top_eagle4_drafter_lb_0_0_synth_1, about 2 h). G8OWN netlist strap census on its funcsim netlist (1,221,265,637 B): PASS. See the row details below the table |
buildxd00 prj.runs/top_eagle4_drafter_lb_0_0_synth_1/top_eagle4_drafter_lb_0_0_sim_netlist.v |
| 02:46:50 | 0x1361 drafter OOC synth run complete (__synthesis_is_complete__, dcp 361,615,502 B); the netlist moved to prj.cache/prj.gen at the same 1,221,265,637 B I censused |
buildxd00 prj.runs |
| 02:5x | 0x1362 drafter OOC synth done; G8OWN funcsim census PASS, identical counts to 0x1361 | build00 |
| 03:48 | 0x1361 top synth_1 done; impl_1 opt done about 04:45 | buildxd00 prj.runs |
| ~06:56 | 0x1361 place_design done (top_wrapper_placed.dcp). Post-place utilization (place_report_utilization_0.rpt) vs parent 0x134D (buildxd02 same report): |
- LUT 1,172,516 (45.55%) vs 1,152,735 (44.78%), +19,781;
- FF 1,018,982 (19.79%) vs 1,017,067, +1,915;
- SLICE 286,981 (89.19%) vs 290,621 (90.32%), -3,640 (placement packing differs);
- BRAM 1,035.5 (27.68%), unchanged;
- URAM 512 (26.60%), unchanged;
- DSP 7,504 (69.17%).
- The G8_OWN cost is about +20K LUT and +2K FF: own_link 2,256 cells, g_own 3,237 cells incl. the 40 RAM64M8 mirror, plus the f_mine/need logic | buildxd00 impl_1 |
| ~06:5x | 0x1362 in Phase 3 Global Placement (
place_design -directive Explore+ subdirectives) | build00 impl_1 |
In-build checks (report-only)
| check | 0x1361 | 0x1362 |
|---|---|---|
| STEP0.5 staged drafter = 845ca21c + PROBE_REV only | PASS (9f4998cc; diff = 1 PROBE_REV line; config diff = PIPE_ROLE 2'd2 + PERSIST 1 only; MLPV_G8_OWN_V 1 x1) | PASS (35327c33, 2-file delta) |
| STEP0.6 derived pkg tcl add_files has mlpv_g8_own_link.sv; no OWN_RED | PASS (:479) | PASS (same derivation) |
| STEP1 component.xml MLPV_G8_OWN | 1 (MODELPARAM + PARAM) | 1 |
| must_pass --pkg --kind MLPV (A1/A3) | PASS (dual 0 and 1) | PASS |
| STEP3 staged-IP list incl. own_link | PASS (no FAIL row) | PASS |
| place recipe | ForceSpreading SUBDIRECTIVE + phys_opt/route AggressiveExplore | same + place -directive Explore: place-arm-Explore-stage PASS |
| BD synth wrapper literal | .MLPV_G8_OWN(1) :573, .S1_OVERLAP(0) |
.MLPV_G8_OWN(1), .S1_OVERLAP(0) |
| netlist census (funcsim) | PASS (see below) | PASS (1,221,263,588 B: u_g8own 2,318, strap-0 arms 0, own_link module x1, f_mine 512, mir 40 RAM64M8) |
B3/B4 (post-synth drafter dcp, scripts/4c1/hedge/g8own_b3b4_census.tcl) |
B4 PASS (u_g8own 2,256 leaf cells; g_own 3,237). The run stopped at the f_mine row on Common 17-736 (collection display limit 500 < 512); the tcl is fixed, and the full re-run was on 0x1362, which has the same RTL bytes. g8own_1361/B3B4_1361_drafter_synth_dcp_partial.txt |
G8OWN_B3B4 PASS: B4 u_g8own 2,256 / g_own 3,237 / f_mine 512 slots / mir 40 RAM64M8 / u_vs 13,470; B3 rank, fin, item_start, tp_m_valid, vs_tx/rx_busy, kcq[0], rd_ok and quiet all driven by FDRE/LUT, none GND/VCC. g8own_1361/B3B4_1362_drafter_synth_dcp.txt |
0x1361 funcsim netlist census details (02:3xZ):
g_g8own.u_g8own2,318 lines;- strap-0 arms
g_g8own_off/g_g8own_tie: 0; - module
..._mlpv_g8_own_linkdefined x1; f_mine_regdistinct slots 512 (= G8_K_C);mir_reg1,412 lines, 40 RAM64M8. The mirror is 512 x 32 b; each RAM64M8 is 64 deep x 7 bits, so it takes 8 x ceil(32/7) = 8 x 5 = 40 cells, the count expected for a LUTRAM mirror;rx_fullnames: 0 (INFO: a comb wire absorbed into a LUT; the EXACT_END source tooth passed pre-launch).
Route interim timing (11:3xZ) and the closure sibling
Both builds are in route rip-up with negative interim timing. Both routes are still running, so nothing is concluded yet.
| build | interim (Route 35-416) | SLL demand per cut (Phase 3.1, demand / ratio) | congestion level |
|---|---|---|---|
| 0x1361 | WNS -1.182 / TNS -976.0 / WHS -0.383 | SLR0<->1 18,870 / 1.122 (0->1 1.150), SLR1<->2 1.003 | global/short 5, timing 7 |
| 0x1362 | WNS -1.101 / TNS -499.1 / WHS -0.001 | SLR0<->1 1.011, SLR1<->2 18,870 / 1.139 (1<-2 1.152) | global/short 6, timing 7 |
| parent 0x134D (closed +0.042 / +0.010) | first -0.909 / -186.9, recovered to 0.000 / 0.000 | SLR0<->1 15,540 / 1.015, SLR1<->2 13,690 / 1.002 | timing 6 |
- Mechanism:
- The parent needed about 1.0x of SLL capacity on both cuts.
- With G8_OWN, the busiest cut needs 1.12-1.15x. The cut moves with the placer: SLR0<->1 for the ForceSpreading recipe (0x1361), SLR1<->2 for Explore + ForceSpreading (0x1362).
- So the G8_OWN logic adds inter-die nets. The own_link sits beside u_vs/TP, the score mirror and f_mine sit in the frontend stripe, and the two are in different SLRs.
- The congestion is at the SLL crossings, not inside a die.
- Sibling recipe:
SSI_SpreadLogic_highandAltSpreadLogic_highare not supported on Versal 2025.1: they revert to Default with WARNING Vivado_Tcl 4-2308. This was measured on the 0x1144R arm B2 and is noted atsibw_fan_impl_reroot.tcl:8.- The legal SLR lever that neither build has tried is the EXB set:
place -directive Explore -subdirective {Floorplan.BalancedSLR.high GPlace.ReduceCongestion.high DPlace.ReducePinDensity.high}, then phys_opt / route / post-route AggressiveExplore (as D7_EXB / C9_EXB / CA_EXB).
- Kit:
scripts/g8ownfix_20261112/(deploy_g8own_ex.sh / run_arm_g8own.sh / push_fw_g8own.sh).- It is a copy of the D7 kit; only the arm root
lutg_g8own_fan_20261112, the log dir and the names change. - The arm starts from 0x1361's own
impl_1/top_wrapper_opt.dcpand hooks. - FPT uses 0x1361's own amc.elf/fpt.bin (push_fw_g8own.sh) plus the AMC_IN_BAND_DBG count row.
- FINAL_TRUE bar: +0.030 setup / +0.010 hold, report-only.
- It is a copy of the D7 kit; only the arm root
- Ids: the arm is the SAME netlist as 0x1361, so reg0x100 = 0xE4B71161 and BUILD_ID = 0xE4B71361. Identity is the PDI md5 plus the tag
amd_v80_gen5x8_25.1_0xE4B71361_G8OWN63.pdi.- A distinct id 0x1363 would need a re-synth, so it is not used. Arm name
G8OWN63.
- A distinct id 0x1363 would need a re-synth, so it is not used. Arm name
- Launch: BLOCKED.
deploy_g8own_ex.sh G8OWN63 buildx04 buildxd00 s3LCg8own61 /shared/build/4c1tp4_1361_g8own_mlpv_stage 0xE4B71361 0xE4B71161 Explore Floorplan.BalancedSLR.high,GPlace.ReduceCongestion.high,DPlace.ReducePinDensity.highwas denied by this session's permission classifier ('Remote Shell Writes' on buildx04, which the original brief had excluded).- Launching it needs the user's permission. build00 / buildxd00 are not eligible either: about 60 GB MemAvailable each, under the 90 GB rule.
Route outcome (14:5xZ)
- 0x1361 route_design completed (
Route 35-16 Router Completed Successfully).- Post Routing Timing Summary: WNS +0.000 (0.000177 ns) / TNS 0.000 / WHS +0.010 / THS 0.000.
- Interim history: -1.182 at rip-up iteration 0, then -0.339, -0.184, -0.049, -0.020, -0.006, then 0.000.
- Post-route phys_opt (AggressiveExplore) runs next. True all-clock timing and the closure re-route come after STEP5.
- 0x1362 went -1.101, -0.606, -0.354, -0.139, -0.035 (Ph8), -0.001 (Ph13), and is at 0.000 / WHS +0.010 inside post-route phys_opt (Physopt 32-1494).
- Both recovered the same way the parent did (0x134D: -0.909, then 0.000). The SLL over-subscription did not stop closure at the run's own bar.
- Both sit at ~0 setup margin, which is below the FINAL_TRUE bar (+0.030 / +0.010). The post_build_closure margin re-route (arm B: setup +0.040 / hold +0.020) is the step that has to lift it.
Result: 0xE4B71361 (buildxd00 s3LCg8own61) -- PDI PRODUCED 16:35:48Z
PDI
- Final FPT PDI (the build's own):
buildxd00:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_tail_asgard_0xE4B71361_66m_626f8c0c.pdi.- It is byte-identical to
/shared/build/s3LCg8own61/aved_oneflip_p2s_tail_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi. - md5
626f8c0c6a0bd2b90ba09a8093a18d43, 90,692,448 B, magic16a5f792.
- It is byte-identical to
- AMC_IN_BAND_DBG: pdi 32 = amc.elf 32 (
build/amc.elf, 2,033,984 B;fpt.bin416 B;_nofpt.pdi90,659,680 B). Equal and non-zero: PASS.
Timing
- Route:
Post Routing Timing Summarygives WNS +0.000 (0.000177 ns), TNS 0, WHS +0.010, THS 0. Post-route phys_opt kept WNS 0.000 / WHS 0.010 (Vivado_Tcl 4-2280). - True timing, all clocks (
impl_1/post_route_phys_opt_report_timing_summary_0.rpt):- Design WNS 0.000, TNS 0.000, 0 failing of 3,390,009 endpoints. WHS +0.010, THS 0.000, 0 failing of 3,389,049. WPWS 0.000.
- This lineage has no margin hook (as for 0x134D/0x1360), so these routed numbers are the true constraints.
- Worst per clock (setup / hold):
clock_reset_clk_usr_1(array fast clock) 0.000 / +0.011;ch0_rxoutclk_30.000 / +0.028;ch0_txoutclk_5+0.002 / +0.022;clk_pl_2+0.003 / +0.012;ch0_rxoutclk_4+0.003 / +0.027;clkout1_primitive+0.004 / +0.010 (worst hold);clock_reset_clk_usr_0+0.023 / +0.011;clk_pl_0+1.264 / +0.026;clk_pl_1+26.804 / +0.021.- Inter-clock: usr_1->usr_0 +0.256 / +0.013, usr_0->usr_1 +0.175 / +0.012, clk_pl_0->clk_pl_1 +1.624 / +2.756.
- async_default worst clk_pl_2 +0.295 / +0.147.
- Every clock is >= 0 for setup and hold, so the PDI is not negative. Setup margin is about 0 (0.000177 ns): the same class as the adjudicated-GO 0x1360 (+0.001 / +0.010 true), and below the +0.030 FINAL_TRUE bar used by the closure siblings.
Closure
post_build_closure.shstarted 16:35Z: arm B, margin re-route setup +0.040 / hold +0.020, then FB with margin off and FINAL_TRUE. It writesimpl_1/closure_g11lsi/FB/fpt/amd_v80_gen5x8_25.1.pdi.
Gates (g8own_1361/GATE_VERDICTS_0x1361_s3LCg8own61.txt, 242 lines, 38 FAIL rows)
- The FAIL set equals the parent 0x134D set after normalising slot and numbers:
- Same ENGCHECK/RESIDENCYCHECK/KVCLRCHECK/tpsafe/CP1Q_LINK/PA/DUALCLK-wrapper/v12 u_lmdown g_wbank/preflash-qwen rows.
- The CONFIG MANIFEST MISMATCH is the same single item,
g_combine_perbank(HEAD_PAR=1 attention branch; it is absent on the MLPV card by design), in both builds. - One extra row is the provisional post-BD copy of that same manifest row. The parent's
strap-watcherFAIL is absent here. - No FAIL row is G8_OWN-caused.
- G8OWN in-build rows, both runs (02:44Z watcher and 16:22Z --final): all PASS.
- packaged config MLPV_G8_OWN_V 1 x1, 0 OWN_RED;
- wrapper literal 1;
- netlist g_g8own.u_g8own lines 784, own_link module x1, strap-0 arms 0;
- f_mine 512 slots;
- mir_reg 1,212 lines;
- rx_full INFO 0.
Flash adjudication 0xE4B71361 (box1-two-v2, 2026-11-12 16:48Z): GO, with a deliberate setup-bar exception
- Final FPT PDI md5 626f8c0c6a0bd2b90ba09a8093a18d43, 90,692,448 B, magic 16a5f792; AMC_IN_BAND_DBG 32 = amc.elf 32.
- True timing on all clocks: setup +0.000177 ns, hold +0.010 ns. 0 failing endpoints.
- Gate FAIL set is identical to the parent 0x134D's. All G8OWN marker rows PASS.
- Exception: setup is below the +0.030 FINAL_TRUE release bar used for the closure siblings today.
- Accepted on precedent: 0x1360, the record MLPV, was adjudicated GO 2026-11-11 at +0.001 / +0.010 true, in the same no-margin-hook lineage.
- Hold meets the +0.010 bar. That differs from CA_KVF, which was NO-GO at hold +0.004.
- Condition: any silicon anomaly (hang, watchdog abort, text mismatch, gate FAIL) quarantines 0x1361, and 0x1360 is restored in the same window.
- Pending M1 preflight: kc_cfg 0x150 must read equal on both MLPV cards.
- Window: talon2 lock OWNER box1-two-v2-g8own 16:49:03Z. Flash 50/63 only; ATT cards stay on 0x13BC.
M1361: m134DB margin-rescue recipe on 0x1361 (coordinator request, 18:2xZ)
- Recipe: copied READ-ONLY from
buildx04:/shared/build/hedge_rescue/kit: rescue_margin.tcl 5d41c5ed, rescue_margin_run.sh e1cd151f, rescue_pkg.sh 33f514f6.- This is the exact kit that produced m134DB 72c697fc. Note the repo copy
scripts/4c1/hedge/rescue/rescue_margin.tclis e87d57f2, a different version. - Archived in
g8own_1361/m134DB_kit/with the m134DB marker and its ^RESCUE lines. - m134DB per pass (true numbers): p0 src +0.001/+0.010; p1 physopt +0.005/+0.010; rip-up 224 nets; p2 route +0.041/+0.019; p3 +0.042/+0.019; FINAL_TRUE +0.042/+0.019.
- This is the exact kit that produced m134DB 72c697fc. Note the repo copy
- Arm:
buildxd00:/shared/build/hedge_rescue_g8own/M1361.- Source: a copy of 0x1361's own
impl_1/top_wrapper_postroute_physopt.dcp, md5 241c998d2a09f391da696ac89ed71259 (source and copy equal). - FW: 0x1361's own
build/amc.elfa46d4d46,fpt.bin45f47718,fpt/fpt_pdi_gen.pya63119fa. - Same netlist, so the same ids: reg0x100 0xE4B71161 / BUILD_ID 0xE4B71361. Identity is the PDI md5.
- Source: a copy of 0x1361's own
- Driver:
scripts/g8ownfix_20261112/m1361_rescue_driver.sh, launched 2026-11-12T18:31:11Z (pid 2638794). Steps:- Wait until the host qualifies (0 other Vivado, MemAvailable >= 90 GB). At launch: 74 GB and 1 Vivado, the closure arm B of 0x1361 still running.
- Run ARM=B MS 0.040 MH 0.012.
- rescue_pkg.sh with the 0x1361 FW.
- Report-only rows: FINAL_TRUE bar +0.030/+0.010, magic, AMC count.
- Write the QUARANTINED marker.
- VM choice: host_build_count_v3 at 18:2xZ found no qualifying 125 GB host (buildxd01-05 and buildzd00-05 each have 1-2 builds or <90 GB; buildxd03 has 0 builds but 65 GB).
- build00 is reserved for 0xE4B71365, and buildx04 is launching 0xE4B713BE.
- So M1361 queues on buildxd00 behind closure arm B.
- Closure arm B status (0x1361, closure_g11lsi):
- s1 phys_opt AggressiveExplore (margin on): true WNS +0.016 / WHS +0.010 (17:42Z).
- s2 route Explore -tns_cleanup: it1_routed.dcp written 18:25Z, measurement and FB pending.
- This is already above 0x134D's closure_g11lsi result (+0.005) but below the m134DB +0.042, so M1361 is the stronger candidate.
Result: 0xE4B71362 (build00 s3LCg8own62, place -directive Explore hedge) -- PDI PRODUCED 18:54:34Z
PDI
- Final FPT PDI:
build00:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_tail_asgard_0xE4B71362_66m_4de5a46a.pdi.- It is byte-identical to the clone's
amd_v80_gen5x8_25.1.pdi. - md5
4de5a46a79895c4d9e3292a272fc345c, 91,488,272 B, magic16a5f792.
- It is byte-identical to the clone's
- AMC_IN_BAND_DBG: pdi 32 = amc.elf 32: PASS. The build compiles its own amc.elf, md5 0d0b3059 (0x1361's is a46d4d46; each build stamps its own FW); fpt.bin 45f47718 is the same.
Timing (post_route_phys_opt_report_timing_summary_0.rpt, no margin hook = true)
- Design WNS 0.000, 0 failing of 3,389,625; WHS +0.010, 0 failing of 3,388,665.
- Per clock (setup / hold):
clock_reset_clk_usr_10.000 / +0.010;clock_reset_clk_usr_0+0.001 / +0.010;- ch0_rxoutclk_3 +0.003 / +0.014; ch0_rxoutclk_2 +0.005 / +0.010;
clkout1_primitive+0.007 / +0.013;- ch0_txoutclk_4 +0.019 / +0.038;
clk_pl_2+0.119 / +0.015;clk_pl_0+1.855 / +0.012;- clk_pl_1 +26.912 / +0.049.
- Inter-clock: usr1->usr0 +0.072 / +0.022; usr0->usr1 +0.060 / +0.026; clk_pl_0->clk_pl_1 +3.291 / +1.722.
- All clocks >= 0. Same ~0 setup margin class as 0x1361. The worst clock is the same,
clock_reset_clk_usr_1, the array fast clock.
Gates (g8own_1361/GATE_VERDICTS_0x1362_s3LCg8own62.txt, 55 FAIL rows vs 34 on 0x1361)
- The extra rows are instrument artifacts, not defects.
- (a) The 02:56Z WATCHER pass of check_v12/enorm censused the funcsim netlist while Vivado was still writing it. It found mlpv_g8_own_link module defs 0, f_mine 0, lanes, e3 pairs and LCACHE absent.
- The --final pass at 18:29Z on the complete netlist shows every G8OWN row PASS: u_g8own 784, module x1, f_mine 512, mir 1,212. That matches 0x1361 exactly.
- It also shows V12 FAIL(1) (the same inherited u_lmdown row) and enorm-d0-engine PASS.
- My own independent census at 02:5xZ waited for the file size to be stable (1,221,263,588 B) and also PASSed.
- (b)
place-spread-implreads 0 because it grepsplace_design -subdirective {..}literally, and 0x1362 writesplace_design -directive Explore -subdirective {..}.- My
place-arm-Explore-implrow PASSes. That is the intended recipe change.
- My
- Everything else is the parent 0x134D set.
- Gate-shape note: the watcher row can PASS-or-FAIL on a half-written netlist. It failed closed here. It should wait for a stable file size; it is report-only, so I am noting it, not editing it.
- post_build_closure for 0x1362 started 18:54Z on build00.
0x1361 closure re-route (post_build_closure.sh, closure_g11lsi arm B) -- PDI 19:36:43Z
- PDI:
buildxd00:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_tail_asgard_s3LCg8own61_MLPV_G11LSI_0xE4B71342_66m_reroute_4f3f4b22.pdi.- It is byte-identical to
impl_1/closure_g11lsi/FB/fpt/amd_v80_gen5x8_25.1.pdi. - md5
4f3f4b229764cce73d278c3795777bdf, 90,692,832 B, magic 16a5f792. - AMC_IN_BAND_DBG 32 = 0x1361's amc.elf 32 (f.bif packs
build/amc.elf).
- It is byte-identical to
- ⚠️ NAME TRAP: the file name says
0xE4B71342because that label is hard-coded in the inherited closure script. The netlist is 0x1361: the source is s3LCg8own61's postroute_physopt dcp, and the marker says 'same netlist as the pipeline PDI'.- On silicon it reads reg0x100 0xE4B71161 / BUILD_ID 0xE4B71361. Identify it by md5, never by file name.
- Steps (true / margined):
- ctl reproduces the build PDI byte-exact (626f8c0c).
- s1 phys_opt AggressiveExplore: margin -0.024 / -0.010, true +0.016 / +0.010.
- s2 route Explore -tns_cleanup: margin -0.024 / -0.010, true +0.016 / +0.010 (no gain).
- FB margin off: FINAL_TRUE WNS +0.016 / WHS +0.010, 0 failing of 3,390,009 / 3,389,049 (
FB/final_true_timing_summary.rpt). - Per clock: usr_1 +0.016 / +0.011; usr_0 +0.051 / +0.011; clkout1 +0.070 / +0.010; clk_pl_2 +0.042 / +0.012; GT ch0_* +0.040..+0.050 / +0.019..+0.033.
- Verdict: better than the build's own +0.000177 and better than 0x134D's closure_g11lsi (+0.005). Still below the +0.030 FINAL_TRUE bar, so M1361 (m134DB recipe) is next on buildxd00.
M1361 adjudication (2026-11-13): GO as a margin-only replacement of the record MLPV 0x1361 on 50/63
The record set is 0x13BC x4 + 0x1361 (626f8c0c), which runs on silicon at 902.77 tok/s @ 6.91 (coordinator). M1361 changes placement and routing only.
Run (buildxd00 /shared/build/hedge_rescue_g8own/M1361)
- Queued 18:31:11Z; started 19:41:12Z (the host qualified: 113 GB, 0 other Vivado); WDI 05:06:15Z; packaged 05:06:44Z.
- Kit: m134DB, rescue_margin.tcl 5d41c5ed, ARM=B, user uncertainty setup 0.040 / hold 0.012 on 49 clocks.
- Evidence:
g8own_1361/m1361/(DRIVER.log, GATE_VERDICTS_M1361.txt, marker, route status, SLR util, FINAL_TRUE summary tables, preflash outputs).
| pass | margined WNS / WHS | nfail setup / hold | TRUE WNS / WHS |
|---|---|---|---|
| p0 src (0x1361 postroute_physopt dcp 241c998d) | margin off | 0 / 0 | +0.000 / +0.010 |
| p0 margin on | -0.040 / -0.002 | 4,701 / 5 | +0.000 / +0.010 |
| p1 phys_opt AggressiveExplore | -0.024 / -0.002 | 2,142 / 2 | +0.016 / +0.010 |
| rip-up worst NFAIL paths | 259 nets | ||
| p2 route AggressiveExplore | +0.001 / +0.004 | 0 / 0 | +0.041 / +0.016 |
| p3 phys_opt + hold_fix + route | +0.001 / +0.004 | 0 / 0 | +0.041 / +0.016 |
| FINAL_TRUE (margin 0) | 0 / 0 | +0.041 / +0.016 |
- For comparison: m134DB on 0x134D gave +0.042 / +0.019; closure_g11lsi on 0x1361 gave +0.016 / +0.010; 0x1361's own build gave +0.000177 / +0.010.
1. True timing, every clock (M1361_FINAL_TRUE.timing_summary.rpt, margin 0)
- Zeroing user uncertainty equals the real constraints here: neither the build's own constraints nor prj.srcs contain any
set_clock_uncertainty, and p0_src reproduces the build's routed numbers. - Design: WNS +0.041, TNS 0, 0 failing of 3,390,009. WHS +0.016, THS 0, 0 failing of 3,389,049. WPWS 0.000 (0 failing).
- Per clock (setup / hold):
clock_reset_clk_usr_1+0.041 / +0.016 (the worst; the array fast clock);clock_reset_clk_usr_0+0.150 / +0.023;clkout1_primitive+0.095 / +0.022;clk_pl_0+0.558 / +0.024;clk_pl_1+26.804 / +0.090;clk_pl_2+0.093 / +0.024;- GT ch0_rx/txoutclk_2..5: setup +0.041..+0.059, hold +0.022..+0.033.
- Inter-clock: usr1->usr0 +0.081 / +0.035; usr0->usr1 +0.193 / +0.053.
- async_default worst: clk_pl_2 +0.295 / +0.147; GT +0.941..+1.525 / +0.199..+0.306.
- Route status: 2,588,691 / 2,588,691 routable nets fully routed, 0 routing errors, 0 unrouted or conflicts at p1/p2/p3/FINAL.
- Clears the +0.030 / +0.010 FINAL_TRUE bar on every clock. 0x1361 itself did not.
2. FPT / AMC
- PDI
amd_v80_gen5x8_25.1.pdimd5e24deb3f6d1de5e5b41964eecc1be7e9, 90,908,832 B, magic16a5f792. - AMC_IN_BAND_DBG 32 = amc.elf 32.
- pdi_combine.bif packs
M1361/out/top_wrapper.pdiwith 0x1361's ownamc.elfa46d4d46 (fpt.bin45f47718,fpt_pdi_gen.pya63119fa). - These are the same FW files as the record 626f8c0c.
- pdi_combine.bif packs
top_wrapper.pdi(90,052,624 B, no AMC) and_nofpt.pdiare NOT flashable (hard rule 11).- box1 copy:
/mnt/data/eagle4_scratch/g8own_silicon_20261112/amd_v80_gen5x8_25.1_0xE4B71361_M1361_e24deb3f.pdi, md5 re-verified e24deb3f, plus.QUARANTINEDand a.GATE_VERDICTS.txtsidecar (0x1361's inherited verdicts + M1361's rows, 246 lines). - preflash_gate.sh --dry-run (box1, expected md5 taken from the buildxd00 rescue_pkg log):
- PDI-1/2/2b/3/4 PASS (86 MB, 32 x AMC).
- Without the ack, exit 1: PDI-5 FAIL, 38 inherited verdict rows.
- With
PREFLASH_ACK_QUARANTINE=1, exit 3: PDI-5 UNVER (acknowledged). - PDI-6/7 UNVER: no strap/efficacy artifact exists. This is the same profile as the adjudicated-GO 0x1360.
- The operator re-runs it on talon2 with the HOST checks live.
3. Marker FLASH-RELEVANT rows and inherited gates
- The M1361 marker and verdicts contain FINAL_TRUE PASS, MAGIC PASS and AMC PASS, and no FLASH-RELEVANT or user-ruling row.
- The inherited 0x1361 build marker (44 lines, 38 FAIL rows;
grep -ci flash-relevant= 0) carries the same set as 0x134D and the adjudicated-GO 0x1360:- gate-selftest SKIP;
- ENG/RESIDENCY/KVCLR checks: no attention on the MLPV card;
- tpsafe;
- v12 u_lmdown;
- pblockgate 17 dead filters;
- BCAST slack < +0.3 and KV_VADDR blind (a calibrated body-cone bar, the same text on 0x134D/0x1360);
- CP1Q_LINK/PA INCONCLUSIVE;
- DUALCLK-wrapper;
- the STEP4d manifest
g_combine_perbank; - qwen_identity_replay_required: Qwen-only, and the flash target is Llama.
- No row is G8_OWN-caused. Every G8OWN netlist row PASSes in both --final runs.
- M1361 does not re-synthesize, so the RTL/netlist gate rows of 0x1361 apply unchanged.
4. Netlist identity to 0x1361
- Lineage, all in 0x1361's impl run on buildxd00 s3LCg8own61: synth_1
top_wrapper.dcp9d594dbf3a568df3f7225d7f7dcc9fe4 -> impl_1top_wrapper_opt.dcp6a446cca191b2b2f2b698d893c9deb98 ->top_wrapper_routed.dcp28e36faa ->top_wrapper_postroute_physopt.dcp241c998d2a09f391da696ac89ed71259. - The PDI 626f8c0c was written from that last dcp. M1361's SRC is a byte-identical copy of it (md5 checked at both ends,
src1361/SRC_MD5). - M1361 then runs only phys_opt (equivalence-preserving physical transforms), unroute/route, and write_device_image. There is no synth, no opt_design re-run and no RTL input.
- The ids are unchanged: PROBE_REV 8'h61 is baked in the netlist, so reg0x100 reads 0xE4B71161 and tp BUILD_ID 0xE4B71361. Silicon identity is the PDI md5.
- The 0x1361 host ABI, boot roots, kc_cfg preflight (audit M1) and v3 xsim evidence apply unchanged.
5. Bookkeeping
scripts/4c1/c4_regmap.pyEXPECT_PDI_MD5[0xE4B71361] = {626f8c0c, e24deb3f}.- Flash record
scripts/4c1/flash_records/flash_7c_tp4_BC_M1361_20261113.json: ATT 0x13BC 41e69bde on 17/3d/bd/e1 unchanged; MLPV 50/63 <- M1361 e24deb3f; card 99 untouched; restore = flash_7c_tp4_BC_1361_20261112.json. - The talon2 path in the record still needs staging (copy plus md5 check).
Verdict
- GO to flash M1361 (e24deb3f) on 50/63 as a margin-only replacement of 0x1361.
- Basis: FINAL_TRUE +0.041 / +0.016 on every clock (0x1361 is +0.000177 / +0.010); magic 16a5f792; AMC 32/32; netlist and ids are those of the record; no new gate row.
- Not flashed by me. First test after flash: register identity (reg100 0xE4B71161, tp 0xE4B71361) plus the record A/B (same kit, d11), expecting tok/s identical to 902.77 within noise.
0x1362 closure re-route (build00, closure_g11lsi arm B) -- PDI 22:13:29Z 11-12
build00:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_tail_asgard_s3LCg8own62_MLPV_G11LSI_0xE4B71342_66m_reroute_4803bdee.pdi.- md5 4803bdee57236cb077509e160edcd169, 91,488,544 B, magic 16a5f792, AMC_IN_BAND_DBG 32.
- FINAL_TRUE +0.016 / +0.010.
- The same hard-coded
0xE4B71342name trap applies: it is the 0x1362 netlist (reg100 0xE4B71162).
- Superseded by M1361 (+0.041 / +0.016 on the record netlist). Not proposed for flash.