RTX 4060 (sm_89) on fa0fd90, 2026-10-16

Repeated after the profile rename (63b7395), on 2648bba, with every result equal once the old profile names are read as the new ones: ../rtx-4060-sm89-renamed/README.md.

Verdict: every step of final_4060.sh exits 0 (run.txt): the preflight (18 of 18), T3 (every per-file run, t3/summary.txt), the four test files of the hopper preset's kernels, the full accumulators of every GPU kernel, T3p (6 passed), the three split-K envelope sweeps (rows equal to those of 2026-10-15), the GPU tier of T4, the bench, and the replay of the H100 PCIe probe through the hopper preset's kernels h4 and h1. The forced h1 replay reports "PASS": false by construction: h1 has no kernel for the 380 mixed-format calls (below).

• Hardware and software: the RTX 4060 host of ../../2026-10-14/rtx4060-host/README.md. Preflight 18 of 18 PASS (preflight.out, preflight.json): driver 570.211.01, Python 3.12.2, torch 2.7.1+cu128, cuBLASLt 12.8.3.14, the 32 packages of the reference lock, fatbins current and equal to the manifest's sha256. It ran that directory's preflight_4060.py with the expected commit set to fa0fd90.
• Code under test: fa0fd90 (a git checkout; mmaemu/, tools/ and tests/ unmodified before and after the run, run.txt). 63b7395, later, renamed the profiles in mmaemu/, tests/ and tools/. Its fatbins are those of ../rc-fatbin/README.md.
• Scripts: final_4060.sh, one step after another, every GPU step under the GPU lock, with the ground truth and the audit data present; final_4060_hk.sh, its step 2b alone (below). Paths: ~/ is the host's home directory.

The same script ran on 7612fd8 (the release-candidate freeze) earlier the same day. The commits between change validated_on of h1 and h4 (sm_89 added), make mmaemu.explain run the planning checks of scaled_mm, and change docstrings, help text and one test (a CPU test, which -m gpu deselects). Every count below equals that run, the sweep outputs are byte-identical, every record of both probe replays is equal apart from its timings, the bench medians differ by at most 0.01 TMAC/s, and python -m mmaemu info now lists h1 and h4 as "validated here".

T3 (t3/)

The counts of 2026-10-15 (../../2026-10-15/rtx-4060-sm89/README.md) file by file, plus the tests of the hopper preset's kernels added since:

• test_kernels_ieee 167 passed (79 then) and test_cuda_runtime 19 passed (13 then): the h1 and h4 tests;
• test_h100_mini (new in tools/run_gpu_tests.sh): 310 passed, its GPU cases (the 124 live H100 calls on the GPU through the reference backend and through auto, which runs h4, and the E4M3 x E4M3 ones through h1); -m gpu deselects its 125 CPU cases;
• test_backend_selection exits 5, nothing collected: it is a CPU-only test file (fake kernels, monkeypatched device) with no GPU-marked test; it runs in T0, and without the marker below.

Full accumulators (t3/di_full_gpu_kernels.log, MMAEMU_DI_FULL=cuda:v1,cuda:v4,cuda:v9,cuda:tc3): 84 passed, the 21 Evo 2 calls of each kernel bit for bit equal to the RTX 4060's recorded accumulators.

The hopper preset's kernels h1 and h4 on this GPU (hopper-kernels/, replay-h4/, replay-h1/)

validated_on of h1 and h4 is {sm_89, sm_120}. sm_120 rests on the runs of ../h1-rtx-5070-ti/README.md and ../h4-rtx-5070-ti/README.md: their four test files and the replay of the H100 probe. sm_89 rests on the same evidence on this GPU:

Check	RTX 5070 Ti (sm_120)	RTX 4060 (sm_89), fa0fd90
tests/test_kernels_ieee.py (no marker filter)	167 passed	167 passed
tests/test_cuda_runtime.py	19 passed	19 passed
tests/test_h100_mini.py	435 passed	435 passed
tests/test_backend_selection.py	26 passed	26 passed
probe replay, --backend auto (h4)	PASS	PASS (replay-h4/mmaemu_replay.json)

• The four test files ran on the package fatbins (MMAEMU_TEST_FATBIN_DIR, as tools/run_gpu_tests.sh sets it), each in its own locked run (hopper-kernels/summary.txt). The first run of this step, inside final_4060.sh before that variable was added to the script, did not set it: the fatbin tests of test_kernels_ieee and test_cuda_runtime skipped (1 passed, 166 skipped; 19 skipped). final_4060_hk.sh ran the step again with it (run.txt, 18:47 to 18:50), and final_4060.sh as committed sets it.
• --backend auto (h4 for E4M3 x E4M3, the reference backend for the other pairs): the 1,043 E4M3 x E4M3 live calls ran on h4 and the 380 mixed-format ones on the reference backend; all 1,423 equal the H100 (1,049 split, 374 unsplit), and the 27 calls the H100 rejected raise ReferenceRejectsError. The 168 Evo 2 outputs ran on h4: 168 of 168 equal. Every live record equals that of the RTX 5070 Ti's replay of the same tree (../rtx-5070-ti-replays/README.md) apart from its timings, and every verdict that of the RTX 5070 Ti's replay of ebd18e1 (../../2026-10-15/h100-probe/results/analysis/mmaemu_replay_cuda_h4.json). The splitk="auto" decisions: 1,353 agree with the H100 and 70 raise SplitKUnknownError, against 1,341 and 82 at ebd18e1; the 12 that changed are now inferred S = 1 by the H100 envelope (source="envelope", verified=True), which ebd18e1 did not have, and agree with the logged decision.
• --backend cuda:h1 (h1 forced for every call): the 1,043 E4M3 x E4M3 live calls ran on h1 and are equal (746 split, 297 unsplit); the 168 Evo 2 outputs ran on h1 and are equal. On the 1,070 E4M3 x E4M3 jobs (27 of them rejected as on the H100) and the 168 Evo 2 outputs every verdict equals that of the RTX 5070 Ti's h1 replay (../../2026-10-15/h100-probe/results/analysis/mmaemu_replay_cuda.json, 3d7dc66). The 380 mixed-format calls raise BackendUnavailableError (h1 implements E4M3 x E4M3 only), so the tool's overall verdict is false; the same 380 calls are equal on the reference backend in the h4 replay. (On the RTX 5070 Ti h1's replay ran through --backend auto before h4 existed, when h1 was the preset's only kernel.)
• Bench exactness (below): h1 equals the CPU reference of the hopper preset on its first 64 rows x 256 columns and h4 equals h1 on every output, at all three shapes.
• python -m mmaemu info (mmaemu_info.txt, from the preflight): cuda:h4 (preset hopper): available via fatbin, validated here, and the same for h1.

T3p (t3p_profile_device.log, 6 passed)

The six tests of 2026-10-15: live torch._scaled_mm parity with mmaemu's reference backend, out= and operand layouts, the profile match, 500 table rows re-queried from cuBLASLt, the live query step, and 11,760 live queries above the bound where mmaemu infers S = 1.

Split-K envelope sweeps (sweeps/)

The three scripts of ../../2026-10-14/splitk-envelope/, unchanged, importing fa0fd90. Their per-query rows (not kept) equal those of 2026-10-15 row for row: 306,859, 663,936 and 666,624 rows; the row files are byte-identical to those of the run on 7612fd8.

T4, GPU tier

../t4/mutants-gpu-rtx4060.json: PASS, 7 kernel mutants killed, the identity control survives.

Bench (bench-*-gauss40.json, .log)

tools/bench_kernels.py --kernels v1,v4,v9,tc3,h1,h4 --shape 1b|7b|40b --M 2048 --data gauss40, every kernel from the release fatbins ("source": "fatbin"). TMAC/s, the median of 5 windows:

Kernel	1B (K 1920, N 5760)	7B (K 4096, N 12288)	40B (K 8192, N 24576)	T6 (docs/performance.md)	Exactness (0 mismatches)
tc3	2.95	2.95	2.94	2.94 / 2.93 / 2.94	every output vs v1
v9	1.54	1.55	1.55	1.53 / 1.54 / 1.54	every output vs v1
h4	1.35	1.41	1.41	(not in T6)	every output vs h1
v4	1.35	1.37	1.37	1.34 / 1.36 / 1.37	every output vs v1
h1	0.69	0.69	0.69	(not in T6)	16,384 outputs vs the CPU reference (hopper)
v1	0.63	0.63	0.63	0.63 / 0.63 / 0.63	16,384 outputs vs the CPU reference (ada)

v1, v4, v9 and tc3 are within 1% of T6.