diff --git a/tools/ca3-v4-amend/pc1-amd-family-20261007.ps1 b/tools/ca3-v4-amend/pc1-amd-family-20261007.ps1 new file mode 100644 index 00000000..45dfed52 --- /dev/null +++ b/tools/ca3-v4-amend/pc1-amd-family-20261007.ps1 @@ -0,0 +1,149 @@ +# 7 October 2026 copy for the PC 1 queue: tools/ca3-pc1-amd/pc1-amd-family.ps1 (run e) unchanged except the kit id, since a machine +# runs a fetch id once and the 0.3.21 host build wipes the jobs folder: fetch-ca3-pc1-amd-family-20261007 (the probe rebuilt on +# build-1 from this tree's proto-opencl/family-probe.c). +# Counter ASIC 3.0, PC 1 AMD job 2 (6 October 2026, run e: the 9070 XT ALONE): item 6's step cost of every reserve candidate family of spec +# 1.13.2 on PC 1's RX 9070 XT (machine ae432dc7, gfx1201), docs/plans/counter-asic-3-reserve.md, status file section +# 3 "Item 6" (the OWED AMD column). Run a (run-ca3-pc1-amd-family-20261006, 150 s, exit 0) measured ordinal 0, PC 1's +# integrated gfx1036 (one CU: alu 40.59 G steps/s against the 9070 XT's expected 1,000 to 2,000): the device list parse +# ran a second -match after the capturing one, which overwrote $Matches, so every device read idx 0 and an empty name, +# and the probe took a bare ordinal. Run b: the probe picks the card BY NAME (`--device-name gfx1201`, the match on the +# newest AMD platform by driver version, the kit worker's dedup rule; `RESULT device_choice name= index= platform= driver= +# cus=`), prints the device's name, CUs and platform on every RESULT line, and refuses with `RESULT error` when no +# gfx1201 is listed; the older-platform duplicate and the gfx1036 run after it by explicit ordinal, as their own labelled +# columns. Run d (17:17Z, the card by name, every variant built) showed a dependent-chain ratio does NOT survive the +# card's own miner: the alu chain read 226 then 195 G steps/s and the ratios swung from 5.9 to 11.8x to 0.17 to 1.3x +# between runs (the loaded card's scheduler, not the ops), while the Mac and 5090 columns were taken with the card alone. +# Run e: the 9070 XT is switched OFF by the RUNNER (`--cards-off amd:gfx1201` at publish, the app's own card path, put +# back on ANY exit with its own enabled flag, identities and cap: rule of 6 October 2026, a script never posts to +# /api/cards) for its three runs; this script only confirms by the process list that the card's opencl worker is gone; +# the 5090 and the 4070 keep mining (NOT --stop-miners). The gfx1036 and the old-platform columns run +# AFTER the restore, beside the miners, as they are not the card under test. mm8 stays exact=unverified: the gfx12 WMMA +# iu8 16x16x16 fragment layout is in no source at hand, and a guessed CPU layout would turn a wrong guess into exact=no. The probe is the family kit's +# family-probe-cl.exe (proto-opencl/family-probe.c of branch ca3-pc1-amd, cross-compiled with mingw, OpenCL.dll at run +# time; the Mac ran the same source bit-exact on Apple OpenCL), three runs per device, each best of 3 with a fresh seed +# per repetition, device event time, bit-exact against the CPU reference on two whole 32-lane groups. Every family is +# tried through every form AMD's OpenCL C offers; a form that does not compile prints build=failed with the first line +# of the build log and the run goes on (run a's finding: AMD's OpenCL C compiles no amd_perm, no sudot4, no WMMA +# builtin; ds_bpermute does compile). Never quits, pauses, resumes or updates the installed app; never writes +# settings.json. Lines: the probe's own `RESULT FAMILY name= ms= gsteps= ratio= exact= path= variant= ... device= cus= +# platform=` and `RESULT FAMILYBEST ...`, each prefixed `RESULT run= dev=