Execution reuse and launch evidence — 2026-10-16¶

This report retains every measured native result, every strict framework control, the one fixed CPU confirmation and all remaining CUDA losses. Fresh API, explicit Session reuse and prepared plans are separate allocation contracts. No universal improvement or complete strict-gate pass follows from selected medians.

The original CPU ABBA retains 3 native strict regressions; its whole-batch verdict is regression. The single fixed CPU confirmation’s pooled verdict is inconclusive, without replacing the original report. CUDA launch native rows: 21 within 5%, 3 inconclusive, 0 regression, 0 comparison failed. In the separately scoped device Session characterization, native point medians are lower than same-artifact Torch in 24/24 rows; every remaining loss and every host/API/prepared row is retained below. Those different-scope point estimates are descriptive, not strict speedup evidence.

Actual measured hardware/compiler identities; full package/driver metadata stays in raw JSON¶
cohort	CPU	platform	native host compiler	host flags	NumPy	Torch	CUDA device	CUDA compiler	CUDA architectures
CPU ABBA	13th Gen Intel(R) Core(TM) i7-13700K	Linux-6.6.87.2-microsoft-standard-WSL2-x86_64-with-glibc2.39	GNU 16.0.1	-O3 -DNDEBUG	2.2.6	2.8.0+cpu	none	none	none
CUDA final scopes	13th Gen Intel(R) Core(TM) i7-13700K	Linux-6.6.87.2-microsoft-standard-WSL2-x86_64-with-glibc2.39	GNU 13.3.0	-O3 -DNDEBUG	2.2.6	2.8.0+cu128	NVIDIA GeForce RTX 2060	NVIDIA 12.8.93	75

Measurement and decision rules¶

The measured host is an i7-13700K under WSL2 with an RTX 2060. CPU and CUDA use separate project-local Python environments. Affinity 0 is a logical CPU constraint, not physical-core isolation. Builds, tests and other project numerical work run outside the quiet campaigns. Windows/display activity, GPU clocks, temperature and other host background work are not controlled. Actual compiler, flags, runtime, device identity and library pools remain in every raw artifact; CPU and CUDA compiler identities are separate.

Inputs are FP64, seed 20261015, rtol=1e-8, atol=1e-12, maxiter=2000, standard CG/PCG with residual_refresh=50. The native mode and dense provider are always explicit. Each process uses three warmups, 30 complete balanced Williams rounds, three repetitions per dense round or one per CSR round. Every invocation charges the initial-vector reset. A CUDA timing block fences before the block and after the last invocation, including host dispatch, iteration/control synchronization and completed GPU work. Input transfers and independent CPU verification downloads are outside timing. These are resident-input solves, not end-to-end host-to-host comparisons.

Strict A/B comparisons preserve full scope/options, fixture hashes and each candidate’s iterations/matvecs. No timing normalization or filtered projection bypasses a work mismatch. Reported A/B medians pool rounds across artifact blocks; brackets show the range of process medians. The saved hierarchical bootstrap resamples artifact blocks, then rounds, 10,000 times with seed 20261015 and 95% percentile intervals. The fixed material-regression budget is 5%: upper ratio bound <=1.05 supports no material regression; lower bound >1.05 signals regression; overlap is inconclusive. This budget does not prove a speedup. Per-case intervals have no simultaneous-coverage correction; serial dependence and few process blocks limit precision. Reference controls can also drift and are retained without subtracting their noise from native results. Numerical range failures are never reclassified as accepted solves.

The three-scope characterization instead reports median-of-process-medians and process min–max ranges. Different estimators are explicitly kept apart. No strict confidence interval is assigned to a different scope, host/device engine or framework implementation. Rounding-induced work differences remain visible as time to independently accepted tolerance. PyTorch runs this project’s eager CG/PCG recurrence, not an official Torch CG solver; Torch’s caching allocator and provider kernels remain part of that candidate.

CPU GEMV/reduction implementation; original ABBA¶

The CPU change targets the unfused (fused=False) GEMV path; fused kernels remain unchanged. All 16 unfused native rows support no material regression with identical iteration/matvec counts. At n=2048, current/baseline medians are 0.579–0.676, or 32–42% lower, across none/Jacobi and prepared/API.

The three original native regressions occur in fused none/API at n=64/256/1024; unchanged NumPy/Torch controls regress in those same cells. All observations remain reported, without causal attribution or subtracting control drift. The original whole-batch verdict remains regression and the fixed pooled confirmation remains inconclusive.

Whole-batch strict verdict: regression. Native rows: 18 within 5%, 11 inconclusive, 3 regression, 0 comparison failed. Every original result remains in this table; a lower point estimate alone is not a passed strict gate.

numpy-out controls: 6 within 5%, 23 inconclusive, 3 regression, 0 comparison failed.

torch-eager-out controls: 6 within 5%, 23 inconclusive, 3 regression, 0 comparison failed.

Every native A/B result¶
configuration	case	candidate	A ms [process range]	B ms [process range]	B/A	95% ratio interval	strict verdict	A iterations/matvecs	B iterations/matvecs	B true residual	threshold
none/prepared/fused=true	dense-64	native	0.0291 [0.0282–0.0301]	0.0301 [0.0298–0.0304]	1.035	[0.956, 1.119]	inconclusive	49/51	49/51	2.072e-06	2.738e-06
none/prepared/fused=true	dense-256	native	0.4288 [0.4246–0.4484]	0.4246 [0.4130–0.4281]	0.990	[0.902, 1.031]	within 5%	67/70	67/70	5.707e-06	6.342e-06
none/prepared/fused=true	dense-1024	native	9.7119 [9.4245–9.8641]	9.5176 [9.4731–9.6779]	0.980	[0.949, 1.037]	within 5%	70/73	70/73	9.505e-06	1.079e-05
none/prepared/fused=true	dense-2048	native	127.1318 [119.2156–134.5177]	133.9246 [126.9155–142.4160]	1.053	[0.929, 1.204]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
none/api/fused=true	dense-64	native	0.0401 [0.0395–0.0405]	0.0597 [0.0590–0.0622]	1.488	[1.268, 1.738]	regression	49/51	49/51	2.072e-06	2.738e-06
none/api/fused=true	dense-256	native	0.4455 [0.4368–0.4600]	0.5573 [0.5491–0.6153]	1.251	[1.159, 1.444]	regression	67/70	67/70	5.707e-06	6.342e-06
none/api/fused=true	dense-1024	native	8.5877 [8.4165–8.7181]	11.6586 [11.5896–11.6967]	1.358	[1.310, 1.420]	regression	70/73	70/73	9.505e-06	1.079e-05
none/api/fused=true	dense-2048	native	141.5111 [128.6857–192.7386]	148.5604 [146.0603–151.5539]	1.050	[0.740, 1.195]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
jacobi/prepared/fused=true	dense-64	native	0.0069 [0.0068–0.0069]	0.0078 [0.0072–0.0100]	1.135	[1.012, 1.734]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/prepared/fused=true	dense-256	native	0.0402 [0.0396–0.0405]	0.0448 [0.0394–0.0491]	1.115	[0.968, 1.250]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/prepared/fused=true	dense-1024	native	1.1747 [1.0770–1.2249]	1.2021 [1.0520–1.4456]	1.023	[0.834, 1.381]	inconclusive	4/6	4/6	1.722e-07	1.079e-05
jacobi/prepared/fused=true	dense-2048	native	10.1645 [9.3780–10.5255]	10.6621 [9.9647–11.6004]	1.049	[0.931, 1.253]	inconclusive	3/5	3/5	1.245e-05	1.474e-05
jacobi/api/fused=true	dense-64	native	0.0136 [0.0136–0.0136]	0.0133 [0.0133–0.0137]	0.981	[0.893, 1.070]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/api/fused=true	dense-256	native	0.0470 [0.0454–0.0479]	0.0503 [0.0487–0.0522]	1.070	[0.977, 1.194]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/api/fused=true	dense-1024	native	0.7500 [0.7497–0.7564]	0.7809 [0.7159–0.8304]	1.041	[0.930, 1.135]	inconclusive	4/6	4/6	1.722e-07	1.079e-05
jacobi/api/fused=true	dense-2048	native	8.9574 [8.1346–9.9082]	8.4909 [8.1763–8.8082]	0.948	[0.793, 1.090]	inconclusive	3/5	3/5	1.245e-05	1.474e-05
none/prepared/fused=false	dense-64	native	0.0594 [0.0588–0.0602]	0.0321 [0.0316–0.0334]	0.541	[0.517, 0.588]	within 5%	49/51	49/51	2.072e-06	2.738e-06
none/prepared/fused=false	dense-256	native	1.4503 [1.4477–1.4532]	0.4429 [0.4417–0.4453]	0.305	[0.295, 0.314]	within 5%	67/70	67/70	5.707e-06	6.342e-06
none/prepared/fused=false	dense-1024	native	30.7270 [30.3756–30.8826]	10.2126 [10.0784–10.4172]	0.332	[0.322, 0.346]	within 5%	70/73	70/73	9.505e-06	1.079e-05
none/prepared/fused=false	dense-2048	native	207.4739 [206.2917–208.1231]	140.2050 [138.7118–141.5010]	0.676	[0.647, 0.712]	within 5%	70/73	70/73	1.323e-05	1.474e-05
none/api/fused=false	dense-64	native	0.0676 [0.0671–0.0683]	0.0433 [0.0408–0.0467]	0.641	[0.575, 0.780]	within 5%	49/51	49/51	2.072e-06	2.738e-06
none/api/fused=false	dense-256	native	1.4852 [1.4782–1.4964]	0.4535 [0.4535–0.4557]	0.305	[0.291, 0.332]	within 5%	67/70	67/70	5.707e-06	6.342e-06
none/api/fused=false	dense-1024	native	30.4321 [30.3481–30.4964]	8.6828 [8.6804–8.8068]	0.285	[0.280, 0.298]	within 5%	70/73	70/73	9.505e-06	1.079e-05
none/api/fused=false	dense-2048	native	215.6052 [213.8326–215.6124]	131.6490 [130.8589–140.3086]	0.611	[0.590, 0.684]	within 5%	70/73	70/73	1.323e-05	1.474e-05
jacobi/prepared/fused=false	dense-64	native	0.0111 [0.0106–0.0115]	0.0079 [0.0078–0.0081]	0.716	[0.643, 0.789]	within 5%	5/7	5/7	1.141e-07	2.738e-06
jacobi/prepared/fused=false	dense-256	native	0.1366 [0.1359–0.1368]	0.0425 [0.0421–0.0431]	0.311	[0.300, 0.328]	within 5%	4/6	4/6	1.099e-06	6.342e-06
jacobi/prepared/fused=false	dense-1024	native	2.7272 [2.5893–2.9558]	1.1113 [1.0821–1.1378]	0.407	[0.359, 0.444]	within 5%	4/6	4/6	1.722e-07	1.079e-05
jacobi/prepared/fused=false	dense-2048	native	17.1767 [15.7812–17.8694]	9.9477 [9.7127–9.9914]	0.579	[0.528, 0.642]	within 5%	3/5	3/5	1.245e-05	1.474e-05
jacobi/api/fused=false	dense-64	native	0.0186 [0.0178–0.0225]	0.0145 [0.0142–0.0152]	0.776	[0.606, 0.877]	within 5%	5/7	5/7	1.141e-07	2.738e-06
jacobi/api/fused=false	dense-256	native	0.1626 [0.1497–0.1731]	0.0548 [0.0502–0.0632]	0.337	[0.287, 0.423]	within 5%	4/6	4/6	1.099e-06	6.342e-06
jacobi/api/fused=false	dense-1024	native	2.8870 [2.8213–2.9104]	0.8120 [0.7768–0.8830]	0.281	[0.259, 0.321]	within 5%	4/6	4/6	1.722e-07	1.079e-05
jacobi/api/fused=false	dense-2048	native	14.6047 [14.1249–15.0368]	8.7879 [8.4863–8.8939]	0.602	[0.551, 0.638]	within 5%	3/5	3/5	1.245e-05	1.474e-05

All framework controls¶
configuration	case	candidate	A ms [process range]	B ms [process range]	B/A	95% ratio interval	strict verdict	A iterations/matvecs	B iterations/matvecs	B true residual	threshold
none/prepared/fused=true	dense-64	numpy-out	0.2070 [0.1978–0.2116]	0.2050 [0.2014–0.2080]	0.990	[0.933, 1.060]	inconclusive	49/51	49/51	2.047e-06	2.738e-06
none/prepared/fused=true	dense-64	torch-eager-out	0.3039 [0.2944–0.3196]	0.3165 [0.3032–0.3272]	1.041	[0.936, 1.166]	inconclusive	49/51	49/51	2.447e-06	2.738e-06
none/prepared/fused=true	dense-256	numpy-out	0.5982 [0.5646–0.6203]	0.6746 [0.6525–0.7000]	1.128	[1.042, 1.250]	inconclusive	67/70	67/70	5.647e-06	6.342e-06
none/prepared/fused=true	dense-256	torch-eager-out	0.7503 [0.7360–0.7915]	0.7385 [0.7210–0.7538]	0.984	[0.870, 1.060]	inconclusive	67/70	67/70	5.726e-06	6.342e-06
none/prepared/fused=true	dense-1024	numpy-out	9.5957 [9.2636–10.1367]	9.5590 [9.5064–9.7983]	0.996	[0.924, 1.066]	inconclusive	70/73	70/73	9.505e-06	1.079e-05
none/prepared/fused=true	dense-1024	torch-eager-out	10.3126 [9.8846–10.6704]	10.1459 [9.8721–10.2397]	0.984	[0.918, 1.073]	inconclusive	70/73	70/73	9.505e-06	1.079e-05
none/prepared/fused=true	dense-2048	numpy-out	99.5133 [88.7630–106.0674]	116.9145 [106.2574–119.9971]	1.175	[0.997, 1.371]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
none/prepared/fused=true	dense-2048	torch-eager-out	114.3616 [105.2146–120.5673]	123.3307 [123.3307–123.5492]	1.078	[0.965, 1.231]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
none/api/fused=true	dense-64	numpy-out	0.2156 [0.2097–0.2191]	0.2995 [0.2868–0.3111]	1.389	[1.287, 1.511]	regression	49/51	49/51	2.047e-06	2.738e-06
none/api/fused=true	dense-64	torch-eager-out	0.3177 [0.3177–0.3193]	0.5164 [0.4869–0.5495]	1.625	[1.396, 1.768]	regression	49/51	49/51	2.447e-06	2.738e-06
none/api/fused=true	dense-256	numpy-out	0.5829 [0.5718–0.5961]	0.9308 [0.9243–0.9365]	1.597	[1.452, 1.741]	regression	67/70	67/70	5.647e-06	6.342e-06
none/api/fused=true	dense-256	torch-eager-out	0.7805 [0.7775–0.8009]	1.4283 [1.2537–1.4588]	1.830	[1.473, 1.932]	regression	67/70	67/70	5.726e-06	6.342e-06
none/api/fused=true	dense-1024	numpy-out	8.3408 [8.3106–8.3671]	11.9604 [11.9437–12.0328]	1.434	[1.405, 1.474]	regression	70/73	70/73	9.505e-06	1.079e-05
none/api/fused=true	dense-1024	torch-eager-out	9.0324 [9.0024–9.0440]	13.6667 [13.3234–13.9336]	1.513	[1.447, 1.572]	regression	70/73	70/73	9.505e-06	1.079e-05
none/api/fused=true	dense-2048	numpy-out	123.2393 [109.5910–170.6077]	126.1648 [118.9971–130.9053]	1.024	[0.685, 1.204]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
none/api/fused=true	dense-2048	torch-eager-out	128.8475 [111.9851–177.8665]	133.2149 [125.2021–135.2453]	1.034	[0.709, 1.226]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
jacobi/prepared/fused=true	dense-64	numpy-out	0.0298 [0.0297–0.0298]	0.0385 [0.0309–0.0419]	1.293	[1.014, 1.466]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/prepared/fused=true	dense-64	torch-eager-out	0.0476 [0.0456–0.0486]	0.0660 [0.0507–0.0755]	1.386	[1.003, 1.689]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/prepared/fused=true	dense-256	numpy-out	0.0562 [0.0525–0.0581]	0.0746 [0.0610–0.0850]	1.327	[1.041, 1.643]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/prepared/fused=true	dense-256	torch-eager-out	0.0720 [0.0681–0.0756]	0.0974 [0.0722–0.1111]	1.354	[0.943, 1.734]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/prepared/fused=true	dense-1024	numpy-out	1.1417 [1.0622–1.1551]	1.1669 [1.0752–1.3057]	1.022	[0.916, 1.243]	inconclusive	4/6	4/6	1.722e-07	1.079e-05
jacobi/prepared/fused=true	dense-1024	torch-eager-out	1.3050 [1.2236–1.4066]	1.3659 [1.2101–1.5161]	1.047	[0.843, 1.275]	inconclusive	4/6	4/6	1.722e-07	1.079e-05
jacobi/prepared/fused=true	dense-2048	numpy-out	8.6770 [7.2163–9.4465]	8.5828 [8.5102–8.6211]	0.989	[0.862, 1.263]	inconclusive	3/5	3/5	1.245e-05	1.474e-05
jacobi/prepared/fused=true	dense-2048	torch-eager-out	9.5979 [9.4643–9.7681]	9.0384 [8.7548–9.1717]	0.942	[0.871, 1.040]	within 5%	3/5	3/5	1.245e-05	1.474e-05
jacobi/api/fused=true	dense-64	numpy-out	0.0405 [0.0403–0.0409]	0.0409 [0.0406–0.0413]	1.009	[0.956, 1.039]	within 5%	5/7	5/7	1.141e-07	2.738e-06
jacobi/api/fused=true	dense-64	torch-eager-out	0.0708 [0.0697–0.0721]	0.0713 [0.0702–0.0721]	1.007	[0.944, 1.062]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/api/fused=true	dense-256	numpy-out	0.0666 [0.0638–0.0686]	0.0642 [0.0621–0.0719]	0.963	[0.890, 1.146]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/api/fused=true	dense-256	torch-eager-out	0.1039 [0.1008–0.1064]	0.1033 [0.1003–0.1097]	0.994	[0.904, 1.134]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/api/fused=true	dense-1024	numpy-out	0.7926 [0.7659–0.8210]	0.8254 [0.7618–0.8591]	1.041	[0.923, 1.138]	inconclusive	4/6	4/6	1.722e-07	1.079e-05
jacobi/api/fused=true	dense-1024	torch-eager-out	0.9153 [0.8671–0.9376]	0.9572 [0.9275–1.0125]	1.046	[0.954, 1.214]	inconclusive	4/6	4/6	1.722e-07	1.079e-05
jacobi/api/fused=true	dense-2048	numpy-out	8.1583 [7.5331–8.5854]	7.2204 [7.0783–7.4492]	0.885	[0.758, 1.001]	within 5%	3/5	3/5	1.245e-05	1.474e-05
jacobi/api/fused=true	dense-2048	torch-eager-out	7.5759 [7.1599–8.3575]	7.7591 [7.6395–7.8702]	1.024	[0.878, 1.129]	inconclusive	3/5	3/5	1.245e-05	1.474e-05
none/prepared/fused=false	dense-64	numpy-out	0.2076 [0.2070–0.2085]	0.2070 [0.2052–0.2090]	0.997	[0.951, 1.027]	within 5%	49/51	49/51	2.047e-06	2.738e-06
none/prepared/fused=false	dense-64	torch-eager-out	0.3270 [0.3266–0.3305]	0.3042 [0.3006–0.3078]	0.930	[0.893, 0.984]	within 5%	49/51	49/51	2.447e-06	2.738e-06
none/prepared/fused=false	dense-256	numpy-out	0.6467 [0.6430–0.6527]	0.6460 [0.5986–0.6982]	0.999	[0.906, 1.109]	inconclusive	67/70	67/70	5.647e-06	6.342e-06
none/prepared/fused=false	dense-256	torch-eager-out	0.7662 [0.7662–0.7665]	0.7931 [0.7752–0.8025]	1.035	[0.971, 1.105]	inconclusive	67/70	67/70	5.726e-06	6.342e-06
none/prepared/fused=false	dense-1024	numpy-out	9.6859 [9.5465–9.7462]	9.9686 [9.5383–10.2351]	1.029	[0.959, 1.094]	inconclusive	70/73	70/73	9.505e-06	1.079e-05
none/prepared/fused=false	dense-1024	torch-eager-out	10.4888 [10.4065–10.7140]	10.6948 [10.5103–11.0233]	1.020	[0.975, 1.081]	inconclusive	70/73	70/73	9.505e-06	1.079e-05
none/prepared/fused=false	dense-2048	numpy-out	95.9973 [91.8661–99.1882]	109.5718 [94.8426–119.9772]	1.141	[0.931, 1.351]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
none/prepared/fused=false	dense-2048	torch-eager-out	111.9841 [110.1135–113.2550]	120.7513 [118.7305–123.6690]	1.078	[1.014, 1.151]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
none/api/fused=false	dense-64	numpy-out	0.2134 [0.2107–0.2146]	0.2050 [0.2010–0.2120]	0.961	[0.922, 1.029]	within 5%	49/51	49/51	2.047e-06	2.738e-06
none/api/fused=false	dense-64	torch-eager-out	0.3175 [0.3131–0.3205]	0.3247 [0.3232–0.3262]	1.023	[0.937, 1.077]	inconclusive	49/51	49/51	2.447e-06	2.738e-06
none/api/fused=false	dense-256	numpy-out	0.6246 [0.5751–0.6536]	0.6426 [0.5955–0.6815]	1.029	[0.895, 1.196]	inconclusive	67/70	67/70	5.647e-06	6.342e-06
none/api/fused=false	dense-256	torch-eager-out	0.8112 [0.7252–0.8365]	0.8149 [0.7859–0.8316]	1.005	[0.935, 1.157]	inconclusive	67/70	67/70	5.726e-06	6.342e-06
none/api/fused=false	dense-1024	numpy-out	8.4386 [8.4287–8.4673]	8.6116 [8.4711–8.8094]	1.021	[0.991, 1.066]	inconclusive	70/73	70/73	9.505e-06	1.079e-05
none/api/fused=false	dense-1024	torch-eager-out	8.9501 [8.9271–9.0104]	9.2984 [9.2866–9.3233]	1.039	[1.008, 1.067]	inconclusive	70/73	70/73	9.505e-06	1.079e-05
none/api/fused=false	dense-2048	numpy-out	118.3176 [115.1932–121.1304]	111.8489 [109.4148–117.0682]	0.945	[0.871, 1.068]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
none/api/fused=false	dense-2048	torch-eager-out	127.0009 [126.6840–128.8439]	116.0280 [112.9043–128.4565]	0.914	[0.861, 1.066]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
jacobi/prepared/fused=false	dense-64	numpy-out	0.0305 [0.0299–0.0328]	0.0311 [0.0305–0.0321]	1.018	[0.898, 1.136]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/prepared/fused=false	dense-64	torch-eager-out	0.0498 [0.0482–0.0523]	0.0514 [0.0510–0.0534]	1.031	[0.927, 1.135]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/prepared/fused=false	dense-256	numpy-out	0.0560 [0.0520–0.0638]	0.0608 [0.0607–0.0613]	1.085	[0.903, 1.227]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/prepared/fused=false	dense-256	torch-eager-out	0.0795 [0.0756–0.0823]	0.0734 [0.0728–0.0736]	0.924	[0.842, 1.027]	within 5%	4/6	4/6	1.099e-06	6.342e-06
jacobi/prepared/fused=false	dense-1024	numpy-out	1.0858 [1.0473–1.1360]	1.0067 [0.9510–1.1347]	0.927	[0.786, 1.124]	inconclusive	4/6	4/6	1.722e-07	1.079e-05
jacobi/prepared/fused=false	dense-1024	torch-eager-out	1.3029 [1.2712–1.3200]	1.1817 [1.1512–1.2294]	0.907	[0.839, 1.039]	within 5%	4/6	4/6	1.722e-07	1.079e-05
jacobi/prepared/fused=false	dense-2048	numpy-out	9.5583 [9.4604–10.0606]	7.5049 [7.3675–7.5702]	0.785	[0.679, 0.849]	within 5%	3/5	3/5	1.245e-05	1.474e-05
jacobi/prepared/fused=false	dense-2048	torch-eager-out	9.7048 [9.2465–9.8699]	8.3512 [7.8973–9.1199]	0.861	[0.761, 1.052]	inconclusive	3/5	3/5	1.245e-05	1.474e-05
jacobi/api/fused=false	dense-64	numpy-out	0.0450 [0.0402–0.0536]	0.0437 [0.0413–0.0466]	0.970	[0.756, 1.188]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/api/fused=false	dense-64	torch-eager-out	0.0884 [0.0770–0.1063]	0.0748 [0.0704–0.0860]	0.846	[0.658, 1.200]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/api/fused=false	dense-256	numpy-out	0.0824 [0.0754–0.0949]	0.0823 [0.0706–0.1040]	0.998	[0.739, 1.401]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/api/fused=false	dense-256	torch-eager-out	0.1448 [0.1166–0.1767]	0.1342 [0.1068–0.1720]	0.927	[0.600, 1.480]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/api/fused=false	dense-1024	numpy-out	0.9938 [0.9002–1.0899]	0.8287 [0.7866–0.8551]	0.834	[0.707, 0.979]	within 5%	4/6	4/6	1.722e-07	1.079e-05
jacobi/api/fused=false	dense-1024	torch-eager-out	1.2149 [1.1030–1.2569]	1.0024 [0.9609–1.0218]	0.825	[0.743, 0.973]	within 5%	4/6	4/6	1.722e-07	1.079e-05
jacobi/api/fused=false	dense-2048	numpy-out	7.3948 [6.6474–8.4181]	7.1452 [6.8059–7.8433]	0.966	[0.786, 1.186]	inconclusive	3/5	3/5	1.245e-05	1.474e-05
jacobi/api/fused=false	dense-2048	torch-eager-out	8.7637 [8.3173–9.3076]	7.6505 [7.4301–7.8587]	0.873	[0.774, 0.976]	within 5%	3/5	3/5	1.245e-05	1.474e-05

Changed same-candidate work rows: 0; no per-iteration normalization.

CPU current native versus matched NumPy/Torch controls¶

Across all 32 current native rows (eight complete configurations, including both fusion flags), 12/32 native medians remain at or above matched NumPy and 8/32 remain at or above matched Torch. Each comparison uses the same configuration, case and current-side artifacts. Medians pool all rounds in the two current artifact blocks, exactly as the A/B tables above; this is not the separate median-of-process-medians estimator. These framework point comparisons have no computed confidence interval or strict speedup verdict. All matched control timings and work counts remain in the preceding tables.

Matched current-side framework summary; ratios are native/reference¶
fusion	reference	native rows	lower native median	native median at/above reference	native/reference range
fused=true	numpy-out	16	11	5	0.147–1.242
fused=true	torch-eager-out	16	12	4	0.095–1.180
fused=false	numpy-out	16	9	7	0.155–1.325
fused=false	torch-eager-out	16	12	4	0.106–1.191
both flags	numpy-out	32	20	12	0.147–1.325
both flags	torch-eager-out	32	24	8	0.095–1.191

Every remaining CPU framework point loss; no candidate/case filtering¶
configuration	case	reference	native ms [process range]	reference ms [process range]	native/reference	higher native median %
none/prepared/fused=true	dense-2048	numpy-out	133.9246 [126.9155–142.4160]	116.9145 [106.2574–119.9971]	1.145	14.55
none/prepared/fused=true	dense-2048	torch-eager-out	133.9246 [126.9155–142.4160]	123.3307 [123.3307–123.5492]	1.086	8.59
none/api/fused=true	dense-2048	numpy-out	148.5604 [146.0603–151.5539]	126.1648 [118.9971–130.9053]	1.178	17.75
none/api/fused=true	dense-2048	torch-eager-out	148.5604 [146.0603–151.5539]	133.2149 [125.2021–135.2453]	1.115	11.52
jacobi/prepared/fused=true	dense-1024	numpy-out	1.2021 [1.0520–1.4456]	1.1669 [1.0752–1.3057]	1.030	3.02
jacobi/prepared/fused=true	dense-2048	numpy-out	10.6621 [9.9647–11.6004]	8.5828 [8.5102–8.6211]	1.242	24.23
jacobi/prepared/fused=true	dense-2048	torch-eager-out	10.6621 [9.9647–11.6004]	9.0384 [8.7548–9.1717]	1.180	17.96
jacobi/api/fused=true	dense-2048	numpy-out	8.4909 [8.1763–8.8082]	7.2204 [7.0783–7.4492]	1.176	17.60
jacobi/api/fused=true	dense-2048	torch-eager-out	8.4909 [8.1763–8.8082]	7.7591 [7.6395–7.8702]	1.094	9.43
none/prepared/fused=false	dense-1024	numpy-out	10.2126 [10.0784–10.4172]	9.9686 [9.5383–10.2351]	1.024	2.45
none/prepared/fused=false	dense-2048	numpy-out	140.2050 [138.7118–141.5010]	109.5718 [94.8426–119.9772]	1.280	27.96
none/prepared/fused=false	dense-2048	torch-eager-out	140.2050 [138.7118–141.5010]	120.7513 [118.7305–123.6690]	1.161	16.11
none/api/fused=false	dense-1024	numpy-out	8.6828 [8.6804–8.8068]	8.6116 [8.4711–8.8094]	1.008	0.83
none/api/fused=false	dense-2048	numpy-out	131.6490 [130.8589–140.3086]	111.8489 [109.4148–117.0682]	1.177	17.70
none/api/fused=false	dense-2048	torch-eager-out	131.6490 [130.8589–140.3086]	116.0280 [112.9043–128.4565]	1.135	13.46
jacobi/prepared/fused=false	dense-1024	numpy-out	1.1113 [1.0821–1.1378]	1.0067 [0.9510–1.1347]	1.104	10.39
jacobi/prepared/fused=false	dense-2048	numpy-out	9.9477 [9.7127–9.9914]	7.5049 [7.3675–7.5702]	1.325	32.55
jacobi/prepared/fused=false	dense-2048	torch-eager-out	9.9477 [9.7127–9.9914]	8.3512 [7.8973–9.1199]	1.191	19.12
jacobi/api/fused=false	dense-2048	numpy-out	8.7879 [8.4863–8.8939]	7.1452 [6.8059–7.8433]	1.230	22.99
jacobi/api/fused=false	dense-2048	torch-eager-out	8.7879 [8.4863–8.8939]	7.6505 [7.4301–7.8587]	1.149	14.87

CPU BLAS was unchanged and not remeasured in this campaign. The historical explicit-BLAS results from 2026-10-15 name their own compiled revision and environment; they do not establish a current all-CPU win or replace the native losses above.

One fixed CPU confirmation; all original observations retained¶

One predeclared unpreconditioned, fused API configuration was measured once more in ABBA+BAAB order: four new artifact blocks per side. The primary pooled comparison includes both original blocks plus all four new blocks per side, across all four sizes and native/NumPy/Torch. The original strict regressions above remain unchanged; this was not an adaptive repeat-until-green campaign.

confirmation-only overall: inconclusive; 5 within 5%, 7 inconclusive, 0 regression, 0 comparison failed.

pooled-original-plus-confirmation overall: inconclusive; 0 within 5%, 12 inconclusive, 0 regression, 0 comparison failed.

All 12 pooled rows; six blocks / 180 rounds per side¶
case	candidate	A ms [process range]	B ms [process range]	B/A	95% ratio interval	strict verdict	A work	B work	B true residual	threshold
dense-64	native	0.0385 [0.0362–0.0405]	0.0444 [0.0374–0.0622]	1.154	[1.017, 1.392]	inconclusive	49/51	49/51	2.072e-06	2.738e-06
dense-64	numpy-out	0.2079 [0.2000–0.2191]	0.2282 [0.2143–0.3111]	1.098	[1.031, 1.370]	inconclusive	49/51	49/51	2.047e-06	2.738e-06
dense-64	torch-eager-out	0.3123 [0.3017–0.3193]	0.3452 [0.3045–0.5495]	1.105	[1.005, 1.542]	inconclusive	49/51	49/51	2.447e-06	2.738e-06
dense-256	native	0.4177 [0.4007–0.4600]	0.4478 [0.4186–0.6153]	1.072	[0.986, 1.292]	inconclusive	67/70	67/70	5.707e-06	6.342e-06
dense-256	numpy-out	0.6049 [0.5498–0.6359]	0.6641 [0.5601–0.9365]	1.098	[0.965, 1.410]	inconclusive	67/70	67/70	5.647e-06	6.342e-06
dense-256	torch-eager-out	0.7793 [0.6819–0.8009]	0.8326 [0.6642–1.4588]	1.068	[0.930, 1.644]	inconclusive	67/70	67/70	5.726e-06	6.342e-06
dense-1024	native	8.0291 [7.7366–8.7181]	8.1066 [7.7409–11.6967]	1.010	[0.940, 1.412]	inconclusive	70/73	70/73	9.505e-06	1.079e-05
dense-1024	numpy-out	8.0161 [7.7588–8.3671]	8.1998 [7.7106–12.0328]	1.023	[0.963, 1.474]	inconclusive	70/73	70/73	9.505e-06	1.079e-05
dense-1024	torch-eager-out	8.5258 [8.1452–9.0440]	8.6422 [8.1308–13.9336]	1.014	[0.948, 1.561]	inconclusive	70/73	70/73	9.505e-06	1.079e-05
dense-2048	native	99.5810 [93.0948–192.7386]	97.5272 [92.5475–151.5539]	0.979	[0.745, 1.445]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
dense-2048	numpy-out	82.0481 [76.6729–170.6077]	81.9235 [76.8896–130.9053]	0.998	[0.732, 1.472]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
dense-2048	torch-eager-out	85.9732 [78.5051–177.8665]	84.4984 [80.4061–135.2453]	0.983	[0.741, 1.509]	inconclusive	70/73	70/73	1.323e-05	1.474e-05

CUDA launch implementation; unchanged prepared/device scopes¶

Whole-batch strict verdict: inconclusive. Native rows: 21 within 5%, 3 inconclusive, 0 regression, 0 comparison failed. Every original result remains in this table; a lower point estimate alone is not a passed strict gate.

torch-cuda-eager controls: 7 within 5%, 9 inconclusive, 0 regression, 0 comparison failed.

Every native A/B result¶
configuration	case	candidate	A ms [process range]	B ms [process range]	B/A	95% ratio interval	strict verdict	A iterations/matvecs	B iterations/matvecs	B true residual	threshold
none/prepared/fused=true/device	dense-64	cuda-native	5.6423 [5.6304–5.6787]	4.3399 [4.1682–4.4648]	0.769	[0.699, 0.886]	within 5%	49/51	49/51	2.045e-06	2.738e-06
none/prepared/fused=true/device	dense-64	cuda-cublas	5.6457 [5.6221–5.7783]	4.7677 [4.7024–5.1161]	0.844	[0.765, 0.960]	within 5%	49/51	49/51	2.206e-06	2.738e-06
none/prepared/fused=true/device	dense-256	cuda-native	7.3222 [7.2508–7.4000]	6.7549 [6.7068–6.7730]	0.923	[0.894, 0.949]	within 5%	67/70	67/70	5.993e-06	6.342e-06
none/prepared/fused=true/device	dense-256	cuda-cublas	7.5940 [7.3911–7.8866]	6.8987 [6.6540–7.0047]	0.908	[0.818, 0.955]	within 5%	67/70	67/70	5.916e-06	6.342e-06
none/prepared/fused=true/device	dense-1024	cuda-native	9.3523 [9.3125–9.3873]	9.0912 [9.0306–9.2297]	0.972	[0.954, 0.999]	within 5%	70/73	70/73	9.505e-06	1.079e-05
none/prepared/fused=true/device	dense-1024	cuda-cublas	8.4293 [8.3157–8.5335]	8.3815 [8.3127–8.4177]	0.994	[0.964, 1.021]	within 5%	70/73	70/73	9.505e-06	1.079e-05
none/prepared/fused=true/device	dense-2048	cuda-native	17.3020 [17.1161–17.6760]	17.3392 [17.3218–17.4454]	1.002	[0.974, 1.027]	within 5%	70/73	70/73	1.323e-05	1.474e-05
none/prepared/fused=true/device	dense-2048	cuda-cublas	15.6735 [15.2615–16.0154]	15.9033 [15.8273–16.0726]	1.015	[0.970, 1.061]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
jacobi/prepared/fused=true/device	dense-64	cuda-native	0.7544 [0.7313–0.7853]	0.6754 [0.6671–0.6892]	0.895	[0.821, 0.966]	within 5%	5/7	5/7	1.141e-07	2.738e-06
jacobi/prepared/fused=true/device	dense-64	cuda-cublas	0.8000 [0.7844–0.8017]	0.7580 [0.7580–0.7619]	0.947	[0.877, 1.057]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/prepared/fused=true/device	dense-256	cuda-native	0.6880 [0.6682–0.7069]	0.6392 [0.6391–0.6409]	0.929	[0.875, 0.988]	within 5%	4/6	4/6	1.099e-06	6.342e-06
jacobi/prepared/fused=true/device	dense-256	cuda-cublas	0.7506 [0.7268–0.7831]	0.6762 [0.6673–0.6909]	0.901	[0.826, 0.973]	within 5%	4/6	4/6	1.099e-06	6.342e-06
jacobi/prepared/fused=true/device	dense-1024	cuda-native	0.9659 [0.9495–0.9812]	0.9085 [0.9071–0.9141]	0.941	[0.909, 0.985]	within 5%	4/6	4/6	1.722e-07	1.079e-05
jacobi/prepared/fused=true/device	dense-1024	cuda-cublas	0.8619 [0.8619–0.8653]	0.7976 [0.7897–0.8019]	0.925	[0.876, 0.982]	within 5%	4/6	4/6	1.722e-07	1.079e-05
jacobi/prepared/fused=true/device	dense-2048	cuda-native	1.2842 [1.2497–1.3012]	1.2212 [1.2129–1.2280]	0.951	[0.906, 0.994]	within 5%	3/5	3/5	1.245e-05	1.474e-05
jacobi/prepared/fused=true/device	dense-2048	cuda-cublas	1.1438 [1.1276–1.1794]	1.1102 [1.0969–1.1191]	0.971	[0.919, 1.045]	within 5%	3/5	3/5	1.245e-05	1.474e-05
none/prepared/fused=true/device	poisson-8	cuda-native	3.0758 [2.6722–3.8269]	2.6011 [2.6011–2.6014]	0.846	[0.646, 1.031]	within 5%	26/28	26/28	2.012e-07	3.055e-07
none/prepared/fused=true/device	poisson-32	cuda-native	12.8803 [12.6022–14.4773]	12.3506 [11.7441–12.6582]	0.959	[0.789, 1.087]	inconclusive	129/133	129/133	1.328e-06	1.367e-06
none/prepared/fused=true/device	poisson-64	cuda-native	30.8441 [29.7714–31.6823]	27.6389 [27.4885–27.6922]	0.896	[0.847, 0.957]	within 5%	290/297	290/297	2.773e-06	2.810e-06
none/prepared/fused=true/device	poisson-128	cuda-native	40.2580 [39.2600–41.3931]	38.9227 [38.3924–39.5243]	0.967	[0.916, 1.013]	within 5%	395/404	395/404	5.661e-06	5.749e-06
jacobi/prepared/fused=true/device	poisson-8	cuda-native	3.0939 [3.0837–3.2870]	2.6334 [2.5714–2.7711]	0.851	[0.719, 0.952]	within 5%	26/28	26/28	2.012e-07	3.055e-07
jacobi/prepared/fused=true/device	poisson-32	cuda-native	16.0244 [15.9783–16.2272]	13.2916 [12.6663–14.4427]	0.829	[0.733, 0.991]	within 5%	129/133	129/133	1.328e-06	1.367e-06
jacobi/prepared/fused=true/device	poisson-64	cuda-native	34.3289 [34.1409–34.5150]	31.6657 [31.2515–32.0494]	0.922	[0.891, 0.956]	within 5%	290/297	290/297	2.773e-06	2.810e-06
jacobi/prepared/fused=true/device	poisson-128	cuda-native	44.8943 [44.0848–46.0453]	42.6465 [42.6465–42.6703]	0.950	[0.908, 0.985]	within 5%	395/404	395/404	5.661e-06	5.749e-06

All framework controls¶
configuration	case	candidate	A ms [process range]	B ms [process range]	B/A	95% ratio interval	strict verdict	A iterations/matvecs	B iterations/matvecs	B true residual	threshold
none/prepared/fused=true/device	dense-64	torch-cuda-eager	13.9748 [12.9417–14.8139]	12.6274 [12.6211–12.6605]	0.904	[0.815, 0.998]	within 5%	49/51	49/51	2.206e-06	2.738e-06
none/prepared/fused=true/device	dense-256	torch-cuda-eager	17.2267 [17.0871–17.3182]	17.1953 [16.9474–17.2825]	0.998	[0.967, 1.027]	within 5%	67/70	67/70	5.651e-06	6.342e-06
none/prepared/fused=true/device	dense-1024	torch-cuda-eager	17.4671 [17.4511–17.5644]	17.8851 [17.7589–18.1346]	1.024	[1.004, 1.044]	within 5%	70/73	70/73	9.505e-06	1.079e-05
none/prepared/fused=true/device	dense-2048	torch-cuda-eager	24.7398 [24.0704–26.0747]	25.7085 [25.6352–25.8954]	1.039	[0.971, 1.079]	inconclusive	70/73	70/73	1.323e-05	1.474e-05
jacobi/prepared/fused=true/device	dense-64	torch-cuda-eager	1.9667 [1.9002–2.0517]	1.8911 [1.8866–1.8972]	0.962	[0.885, 1.075]	inconclusive	5/7	5/7	1.141e-07	2.738e-06
jacobi/prepared/fused=true/device	dense-256	torch-cuda-eager	1.6377 [1.5824–1.7293]	1.5830 [1.5645–1.5957]	0.967	[0.872, 1.056]	inconclusive	4/6	4/6	1.099e-06	6.342e-06
jacobi/prepared/fused=true/device	dense-1024	torch-cuda-eager	1.7921 [1.6220–1.8222]	1.5904 [1.5891–1.6058]	0.887	[0.813, 1.027]	within 5%	4/6	4/6	1.722e-07	1.079e-05
jacobi/prepared/fused=true/device	dense-2048	torch-cuda-eager	1.8595 [1.7699–1.9978]	1.7353 [1.7353–1.7559]	0.933	[0.817, 1.040]	within 5%	3/5	3/5	1.245e-05	1.474e-05
none/prepared/fused=true/device	poisson-8	torch-cuda-eager	7.4530 [6.5675–8.1146]	7.5784 [7.4172–7.6237]	1.017	[0.854, 1.255]	inconclusive	26/28	26/28	2.012e-07	3.055e-07
none/prepared/fused=true/device	poisson-32	torch-cuda-eager	31.1815 [30.3226–32.7041]	36.2604 [34.5338–37.6677]	1.163	[1.045, 1.252]	inconclusive	129/133	129/133	1.328e-06	1.367e-06
none/prepared/fused=true/device	poisson-64	torch-cuda-eager	72.6044 [68.7849–82.4225]	75.7878 [73.0645–77.6817]	1.044	[0.879, 1.135]	inconclusive	290/297	290/297	2.773e-06	2.810e-06
none/prepared/fused=true/device	poisson-128	torch-cuda-eager	95.3272 [94.4449–96.8430]	103.8424 [101.8308–105.5142]	1.089	[1.010, 1.178]	inconclusive	395/404	395/404	5.661e-06	5.749e-06
jacobi/prepared/fused=true/device	poisson-8	torch-cuda-eager	9.3422 [8.8440–10.2706]	8.8418 [8.2586–9.6400]	0.946	[0.775, 1.164]	inconclusive	26/28	26/28	2.012e-07	3.055e-07
jacobi/prepared/fused=true/device	poisson-32	torch-cuda-eager	45.5204 [43.5759–47.6907]	43.5702 [42.2954–45.1413]	0.957	[0.882, 1.038]	within 5%	129/133	129/133	1.328e-06	1.367e-06
jacobi/prepared/fused=true/device	poisson-64	torch-cuda-eager	101.8374 [100.2783–112.3567]	105.1070 [103.4498–109.1952]	1.032	[0.905, 1.106]	inconclusive	290/297	290/297	2.773e-06	2.810e-06
jacobi/prepared/fused=true/device	poisson-128	torch-cuda-eager	133.7335 [133.7335–133.8111]	133.4945 [132.7613–136.5448]	0.998	[0.982, 1.034]	within 5%	395/404	395/404	5.661e-06	5.749e-06

Changed same-candidate work rows: 0; no per-iteration normalization.

CUDA fresh API, explicit Session reuse and prepared plans¶

These are resident-input completed-wall measurements. Fresh API includes a new output, workspace/handle/scratch lifecycle and per-call Jacobi. Session reuses only iterative workspace/resources, allocates a fresh output and recomputes Jacobi per call; it caches neither the operator nor M. Prepared plans cache an owning operator snapshot and M and reuse output/workspace. Torch follows the corresponding allocation contract with its own caching allocator and project-written eager CG recurrence. Its native host/device execution selector is disclosed and does not change Torch’s engine.

Each cell uses two fresh processes in forward/reverse configuration passes. Values are the median of the two process medians, with their min–max range; those ranges are not confidence intervals. Scopes charge different work, so cross-scope and host/device differences are descriptive and are never strict A/B verdicts.

Descriptive same-artifact native/Torch median inventory¶
scope	native mode	candidate	lower native median	rows
api	device	cuda-cublas	4	8
api	device	cuda-native	16	16
api	host	cuda-cublas	4	8
api	host	cuda-native	16	16
prepared	device	cuda-cublas	8	8
prepared	device	cuda-native	16	16
prepared	host	cuda-cublas	8	8
prepared	host	cuda-native	16	16
session	device	cuda-cublas	8	8
session	device	cuda-native	16	16
session	host	cuda-cublas	8	8
session	host	cuda-native	16	16

All api native rows; native/Torch ratio below 1 means a lower native median¶
suite	M	mode	case	native candidate	native ms [process range]	Torch ms [process range]	native/Torch	native work	Torch work	native true residual	Torch true residual	threshold
cuda-dense-f64-v1	none	host	dense-64	cuda-native	6.8317 [6.6240–7.0394]	12.8566 [12.3754–13.3377]	0.531	49/51	49/51	2.045e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	host	dense-64	cuda-cublas	8.9091 [8.6765–9.1417]	12.8566 [12.3754–13.3377]	0.693	49/51	49/51	2.206e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	host	dense-256	cuda-native	9.3219 [9.1807–9.4630]	17.3021 [16.9480–17.6562]	0.539	67/70	67/70	5.993e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	host	dense-256	cuda-cublas	11.0772 [10.9205–11.2339]	17.3021 [16.9480–17.6562]	0.640	67/70	67/70	5.916e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	host	dense-1024	cuda-native	11.8353 [11.7847–11.8859]	17.8950 [17.7652–18.0247]	0.661	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	host	dense-1024	cuda-cublas	12.8860 [12.7964–12.9755]	17.8950 [17.7652–18.0247]	0.720	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	host	dense-2048	cuda-native	19.8583 [19.6171–20.0995]	25.6089 [25.2195–25.9982]	0.775	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	none	host	dense-2048	cuda-cublas	20.4723 [19.9465–20.9982]	25.6089 [25.2195–25.9982]	0.799	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	none	device	dense-64	cuda-native	4.3666 [3.9918–4.7414]	12.4976 [12.2556–12.7395]	0.349	49/51	49/51	2.045e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	device	dense-64	cuda-cublas	6.8429 [6.6823–7.0034]	12.4976 [12.2556–12.7395]	0.548	49/51	49/51	2.206e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	device	dense-256	cuda-native	6.8552 [6.7778–6.9326]	18.8295 [17.9964–19.6627]	0.364	67/70	67/70	5.993e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	device	dense-256	cuda-cublas	9.2182 [9.2176–9.2189]	18.8295 [17.9964–19.6627]	0.490	67/70	67/70	5.916e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	device	dense-1024	cuda-native	9.4237 [9.2222–9.6253]	17.7847 [17.4812–18.0882]	0.530	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	device	dense-1024	cuda-cublas	10.3744 [10.1999–10.5489]	17.7847 [17.4812–18.0882]	0.583	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	device	dense-2048	cuda-native	17.3648 [17.1128–17.6168]	25.2308 [25.2025–25.2591]	0.688	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	none	device	dense-2048	cuda-cublas	18.0229 [17.8412–18.2045]	25.2308 [25.2025–25.2591]	0.714	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	jacobi	host	dense-64	cuda-native	1.1782 [1.1564–1.2000]	2.5677 [2.4499–2.6855]	0.459	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	host	dense-64	cuda-cublas	3.0670 [3.0107–3.1233]	2.5677 [2.4499–2.6855]	1.194	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	host	dense-256	cuda-native	1.0196 [1.0028–1.0364]	2.0143 [1.9079–2.1206]	0.506	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	host	dense-256	cuda-cublas	2.9675 [2.9133–3.0216]	2.0143 [1.9079–2.1206]	1.473	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	host	dense-1024	cuda-native	1.2840 [1.2595–1.3085]	2.2908 [2.2025–2.3792]	0.560	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	host	dense-1024	cuda-cublas	3.1502 [2.9970–3.3034]	2.2908 [2.2025–2.3792]	1.375	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	host	dense-2048	cuda-native	1.5889 [1.5173–1.6604]	2.1715 [2.1256–2.2174]	0.732	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-dense-f64-v1	jacobi	host	dense-2048	cuda-cublas	3.0547 [2.9548–3.1547]	2.1715 [2.1256–2.2174]	1.407	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-dense-f64-v1	jacobi	device	dense-64	cuda-native	0.7616 [0.7389–0.7843]	2.3936 [2.3283–2.4590]	0.318	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	device	dense-64	cuda-cublas	2.6093 [2.5200–2.6986]	2.3936 [2.3283–2.4590]	1.090	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	device	dense-256	cuda-native	0.7471 [0.7234–0.7707]	2.0369 [1.9728–2.1010]	0.367	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	device	dense-256	cuda-cublas	2.6720 [2.6441–2.6998]	2.0369 [1.9728–2.1010]	1.312	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	device	dense-1024	cuda-native	1.0086 [0.9733–1.0440]	2.2537 [2.0287–2.4787]	0.448	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	device	dense-1024	cuda-cublas	2.7173 [2.5266–2.9079]	2.2537 [2.0287–2.4787]	1.206	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	device	dense-2048	cuda-native	1.3276 [1.3063–1.3488]	2.2230 [2.1870–2.2590]	0.597	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-dense-f64-v1	jacobi	device	dense-2048	cuda-cublas	2.9176 [2.8424–2.9929]	2.2230 [2.1870–2.2590]	1.312	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-csr-f64-v1	none	host	poisson-8	cuda-native	3.1153 [3.0455–3.1851]	6.1137 [5.9483–6.2791]	0.510	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	none	host	poisson-32	cuda-native	18.0034 [17.8652–18.1417]	32.3880 [31.4625–33.3135]	0.556	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	none	host	poisson-64	cuda-native	38.0655 [37.5857–38.5452]	73.0056 [72.0612–73.9500]	0.521	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	none	host	poisson-128	cuda-native	54.1683 [53.9058–54.4309]	96.0924 [96.0586–96.1262]	0.564	395/404	395/404	5.661e-06	5.661e-06	5.749e-06
cuda-csr-f64-v1	none	device	poisson-8	cuda-native	2.3929 [2.2283–2.5575]	6.4026 [5.8504–6.9548]	0.374	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	none	device	poisson-32	cuda-native	11.0328 [10.1745–11.8912]	32.5490 [30.5717–34.5263]	0.339	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	none	device	poisson-64	cuda-native	25.9250 [25.8256–26.0244]	71.3727 [69.1815–73.5639]	0.363	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	none	device	poisson-128	cuda-native	35.5576 [35.1683–35.9469]	96.2196 [94.5242–97.9149]	0.370	395/404	395/404	5.661e-06	5.661e-06	5.749e-06
cuda-csr-f64-v1	jacobi	host	poisson-8	cuda-native	4.8662 [4.7789–4.9535]	8.9841 [8.8931–9.0751]	0.542	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	jacobi	host	poisson-32	cuda-native	24.8968 [24.8184–24.9753]	45.3477 [44.9525–45.7429]	0.549	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	jacobi	host	poisson-64	cuda-native	56.1891 [55.8722–56.5060]	101.9474 [100.9209–102.9738]	0.551	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	jacobi	host	poisson-128	cuda-native	79.1395 [78.3049–79.9741]	136.4509 [133.4917–139.4100]	0.580	395/404	395/404	5.661e-06	5.661e-06	5.749e-06
cuda-csr-f64-v1	jacobi	device	poisson-8	cuda-native	2.7127 [2.6226–2.8029]	9.0698 [8.5426–9.5969]	0.299	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	jacobi	device	poisson-32	cuda-native	14.6082 [14.2953–14.9212]	49.7788 [47.2998–52.2578]	0.293	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	jacobi	device	poisson-64	cuda-native	30.3176 [29.5307–31.1045]	100.0418 [97.9815–102.1021]	0.303	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	jacobi	device	poisson-128	cuda-native	42.1653 [40.7842–43.5465]	132.5107 [131.3209–133.7005]	0.318	395/404	395/404	5.661e-06	5.661e-06	5.749e-06

All session native rows; native/Torch ratio below 1 means a lower native median¶
suite	M	mode	case	native candidate	native ms [process range]	Torch ms [process range]	native/Torch	native work	Torch work	native true residual	Torch true residual	threshold
cuda-dense-f64-v1	none	host	dense-64	cuda-native	6.4409 [6.2005–6.6813]	12.9228 [12.4931–13.3526]	0.498	49/51	49/51	2.045e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	host	dense-64	cuda-cublas	6.8715 [6.6620–7.0810]	12.9228 [12.4931–13.3526]	0.532	49/51	49/51	2.206e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	host	dense-256	cuda-native	9.2630 [9.2324–9.2935]	17.2045 [16.7177–17.6913]	0.538	67/70	67/70	5.993e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	host	dense-256	cuda-cublas	9.1076 [9.0866–9.1287]	17.2045 [16.7177–17.6913]	0.529	67/70	67/70	5.916e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	host	dense-1024	cuda-native	12.2839 [11.9368–12.6311]	18.6189 [17.5533–19.6845]	0.660	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	host	dense-1024	cuda-cublas	11.0794 [10.6400–11.5188]	18.6189 [17.5533–19.6845]	0.595	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	host	dense-2048	cuda-native	19.7945 [19.6290–19.9599]	25.1646 [24.9305–25.3987]	0.787	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	none	host	dense-2048	cuda-cublas	18.5692 [18.4049–18.7336]	25.1646 [24.9305–25.3987]	0.738	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	none	device	dense-64	cuda-native	4.4184 [3.9722–4.8645]	12.6469 [12.3479–12.9458]	0.349	49/51	49/51	2.045e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	device	dense-64	cuda-cublas	4.9247 [4.8633–4.9861]	12.6469 [12.3479–12.9458]	0.389	49/51	49/51	2.206e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	device	dense-256	cuda-native	6.6677 [6.2935–7.0419]	17.3292 [17.1379–17.5205]	0.385	67/70	67/70	5.993e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	device	dense-256	cuda-cublas	7.3667 [6.9940–7.7394]	17.3292 [17.1379–17.5205]	0.425	67/70	67/70	5.916e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	device	dense-1024	cuda-native	9.5091 [9.4184–9.5998]	18.2552 [17.9356–18.5747]	0.521	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	device	dense-1024	cuda-cublas	8.8078 [8.7250–8.8906]	18.2552 [17.9356–18.5747]	0.482	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	device	dense-2048	cuda-native	17.3885 [17.2662–17.5109]	25.5453 [25.3888–25.7019]	0.681	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	none	device	dense-2048	cuda-cublas	16.1557 [16.0732–16.2381]	25.5453 [25.3888–25.7019]	0.632	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	jacobi	host	dense-64	cuda-native	1.1015 [1.0751–1.1278]	2.4124 [2.3807–2.4440]	0.457	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	host	dense-64	cuda-cublas	1.1535 [1.0911–1.2160]	2.4124 [2.3807–2.4440]	0.478	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	host	dense-256	cuda-native	1.0062 [1.0010–1.0113]	2.1549 [2.0227–2.2871]	0.467	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	host	dense-256	cuda-cublas	0.9832 [0.9694–0.9971]	2.1549 [2.0227–2.2871]	0.456	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	host	dense-1024	cuda-native	1.3550 [1.3094–1.4006]	2.1750 [2.0980–2.2519]	0.623	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	host	dense-1024	cuda-cublas	1.2267 [1.2114–1.2420]	2.1750 [2.0980–2.2519]	0.564	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	host	dense-2048	cuda-native	1.6103 [1.5925–1.6281]	2.3077 [2.2112–2.4041]	0.698	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-dense-f64-v1	jacobi	host	dense-2048	cuda-cublas	1.4380 [1.3926–1.4835]	2.3077 [2.2112–2.4041]	0.623	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-dense-f64-v1	jacobi	device	dense-64	cuda-native	0.8027 [0.7629–0.8425]	2.2998 [2.1794–2.4201]	0.349	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	device	dense-64	cuda-cublas	0.8598 [0.8466–0.8729]	2.2998 [2.1794–2.4201]	0.374	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	device	dense-256	cuda-native	0.7642 [0.7449–0.7834]	2.0341 [2.0110–2.0572]	0.376	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	device	dense-256	cuda-cublas	0.7844 [0.7717–0.7971]	2.0341 [2.0110–2.0572]	0.386	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	device	dense-1024	cuda-native	1.0651 [1.0558–1.0743]	2.1836 [2.1224–2.2448]	0.488	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	device	dense-1024	cuda-cublas	0.9391 [0.9198–0.9583]	2.1836 [2.1224–2.2448]	0.430	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	device	dense-2048	cuda-native	1.3490 [1.3417–1.3563]	2.2248 [2.1149–2.3347]	0.606	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-dense-f64-v1	jacobi	device	dense-2048	cuda-cublas	1.2367 [1.2078–1.2655]	2.2248 [2.1149–2.3347]	0.556	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-csr-f64-v1	none	host	poisson-8	cuda-native	3.3751 [3.2767–3.4735]	6.8638 [6.8021–6.9255]	0.492	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	none	host	poisson-32	cuda-native	17.0735 [16.5414–17.6056]	29.6516 [28.4126–30.8906]	0.576	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	none	host	poisson-64	cuda-native	38.0268 [37.7248–38.3288]	72.2389 [71.9787–72.4990]	0.526	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	none	host	poisson-128	cuda-native	54.0047 [53.3658–54.6436]	100.3586 [96.7029–104.0142]	0.538	395/404	395/404	5.661e-06	5.661e-06	5.749e-06
cuda-csr-f64-v1	none	device	poisson-8	cuda-native	2.3345 [2.2651–2.4040]	6.2671 [6.0955–6.4386]	0.373	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	none	device	poisson-32	cuda-native	10.5627 [10.4353–10.6901]	31.1188 [30.8150–31.4226]	0.339	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	none	device	poisson-64	cuda-native	25.7859 [25.5527–26.0192]	69.7124 [68.7437–70.6812]	0.370	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	none	device	poisson-128	cuda-native	35.5474 [34.3324–36.7623]	94.5728 [93.5048–95.6409]	0.376	395/404	395/404	5.661e-06	5.661e-06	5.749e-06
cuda-csr-f64-v1	jacobi	host	poisson-8	cuda-native	4.6942 [4.5808–4.8076]	9.2137 [9.0813–9.3462]	0.509	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	jacobi	host	poisson-32	cuda-native	24.7660 [24.5695–24.9625]	45.8044 [44.4537–47.1551]	0.541	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	jacobi	host	poisson-64	cuda-native	56.1782 [54.3404–58.0160]	102.3066 [98.0306–106.5826]	0.549	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	jacobi	host	poisson-128	cuda-native	81.1679 [78.9226–83.4132]	144.2063 [138.6464–149.7662]	0.563	395/404	395/404	5.661e-06	5.661e-06	5.749e-06
cuda-csr-f64-v1	jacobi	device	poisson-8	cuda-native	2.8922 [2.7450–3.0394]	10.5379 [8.4859–12.5899]	0.274	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	jacobi	device	poisson-32	cuda-native	14.8867 [14.7195–15.0539]	47.7860 [46.7500–48.8220]	0.312	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	jacobi	device	poisson-64	cuda-native	34.0727 [32.1407–36.0046]	111.7147 [104.2449–119.1845]	0.305	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	jacobi	device	poisson-128	cuda-native	45.2147 [40.9902–49.4391]	145.6135 [134.4092–156.8178]	0.311	395/404	395/404	5.661e-06	5.661e-06	5.749e-06

All prepared native rows; native/Torch ratio below 1 means a lower native median¶
suite	M	mode	case	native candidate	native ms [process range]	Torch ms [process range]	native/Torch	native work	Torch work	native true residual	Torch true residual	threshold
cuda-dense-f64-v1	none	host	dense-64	cuda-native	6.6642 [6.5761–6.7523]	13.0512 [12.5074–13.5949]	0.511	49/51	49/51	2.045e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	host	dense-64	cuda-cublas	6.8374 [6.6113–7.0636]	13.0512 [12.5074–13.5949]	0.524	49/51	49/51	2.206e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	host	dense-256	cuda-native	9.4742 [9.1256–9.8227]	18.7622 [16.9894–20.5349]	0.505	67/70	67/70	5.993e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	host	dense-256	cuda-cublas	9.3612 [9.0134–9.7090]	18.7622 [16.9894–20.5349]	0.499	67/70	67/70	5.916e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	host	dense-1024	cuda-native	11.9429 [11.6990–12.1868]	18.2972 [17.8519–18.7426]	0.653	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	host	dense-1024	cuda-cublas	10.7081 [10.6041–10.8120]	18.2972 [17.8519–18.7426]	0.585	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	host	dense-2048	cuda-native	19.7652 [19.5814–19.9490]	25.4825 [25.4728–25.4923]	0.776	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	none	host	dense-2048	cuda-cublas	18.2722 [18.2109–18.3335]	25.4825 [25.4728–25.4923]	0.717	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	none	device	dense-64	cuda-native	3.9866 [3.8742–4.0990]	12.7281 [12.3809–13.0753]	0.313	49/51	49/51	2.045e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	device	dense-64	cuda-cublas	4.9011 [4.3832–5.4189]	12.7281 [12.3809–13.0753]	0.385	49/51	49/51	2.206e-06	2.206e-06	2.738e-06
cuda-dense-f64-v1	none	device	dense-256	cuda-native	6.8289 [6.7420–6.9159]	17.9631 [17.5579–18.3684]	0.380	67/70	67/70	5.993e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	device	dense-256	cuda-cublas	7.1670 [7.0710–7.2630]	17.9631 [17.5579–18.3684]	0.399	67/70	67/70	5.916e-06	5.651e-06	6.342e-06
cuda-dense-f64-v1	none	device	dense-1024	cuda-native	9.4275 [9.2828–9.5722]	17.9883 [17.5610–18.4157]	0.524	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	device	dense-1024	cuda-cublas	8.4932 [8.3483–8.6381]	17.9883 [17.5610–18.4157]	0.472	70/73	70/73	9.505e-06	9.505e-06	1.079e-05
cuda-dense-f64-v1	none	device	dense-2048	cuda-native	17.1752 [17.0219–17.3285]	25.5949 [25.5549–25.6349]	0.671	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	none	device	dense-2048	cuda-cublas	15.8856 [15.6489–16.1222]	25.5949 [25.5549–25.6349]	0.621	70/73	70/73	1.323e-05	1.323e-05	1.474e-05
cuda-dense-f64-v1	jacobi	host	dense-64	cuda-native	1.0022 [0.9655–1.0389]	1.9897 [1.9648–2.0146]	0.504	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	host	dense-64	cuda-cublas	0.9923 [0.9783–1.0063]	1.9897 [1.9648–2.0146]	0.499	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	host	dense-256	cuda-native	0.9412 [0.9057–0.9767]	1.8176 [1.6105–2.0246]	0.518	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	host	dense-256	cuda-cublas	0.9759 [0.9211–1.0307]	1.8176 [1.6105–2.0246]	0.537	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	host	dense-1024	cuda-native	1.3118 [1.2929–1.3308]	1.9978 [1.7732–2.2224]	0.657	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	host	dense-1024	cuda-cublas	1.1540 [1.1077–1.2003]	1.9978 [1.7732–2.2224]	0.578	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	host	dense-2048	cuda-native	1.5912 [1.5754–1.6069]	1.9501 [1.8519–2.0483]	0.816	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-dense-f64-v1	jacobi	host	dense-2048	cuda-cublas	1.4282 [1.3612–1.4953]	1.9501 [1.8519–2.0483]	0.732	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-dense-f64-v1	jacobi	device	dense-64	cuda-native	0.7311 [0.7137–0.7485]	1.9476 [1.8765–2.0187]	0.375	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	device	dense-64	cuda-cublas	0.7556 [0.7226–0.7886]	1.9476 [1.8765–2.0187]	0.388	5/7	5/7	1.141e-07	1.141e-07	2.738e-06
cuda-dense-f64-v1	jacobi	device	dense-256	cuda-native	0.6700 [0.6554–0.6847]	1.6443 [1.6069–1.6817]	0.407	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	device	dense-256	cuda-cublas	0.6565 [0.6456–0.6674]	1.6443 [1.6069–1.6817]	0.399	4/6	4/6	1.099e-06	1.099e-06	6.342e-06
cuda-dense-f64-v1	jacobi	device	dense-1024	cuda-native	0.9274 [0.9234–0.9313]	1.6576 [1.6412–1.6740]	0.559	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	device	dense-1024	cuda-cublas	0.8244 [0.8157–0.8331]	1.6576 [1.6412–1.6740]	0.497	4/6	4/6	1.722e-07	1.722e-07	1.079e-05
cuda-dense-f64-v1	jacobi	device	dense-2048	cuda-native	1.2606 [1.2522–1.2690]	1.8898 [1.8479–1.9317]	0.667	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-dense-f64-v1	jacobi	device	dense-2048	cuda-cublas	1.1380 [1.1245–1.1514]	1.8898 [1.8479–1.9317]	0.602	3/5	3/5	1.245e-05	1.245e-05	1.474e-05
cuda-csr-f64-v1	none	host	poisson-8	cuda-native	3.2242 [3.1873–3.2610]	6.1342 [5.6740–6.5945]	0.526	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	none	host	poisson-32	cuda-native	18.0640 [17.7476–18.3805]	32.2566 [31.0885–33.4247]	0.560	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	none	host	poisson-64	cuda-native	37.0925 [37.0580–37.1271]	69.9784 [68.6766–71.2803]	0.530	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	none	host	poisson-128	cuda-native	53.7910 [52.8938–54.6882]	100.4247 [96.1026–104.7467]	0.536	395/404	395/404	5.661e-06	5.661e-06	5.749e-06
cuda-csr-f64-v1	none	device	poisson-8	cuda-native	2.1858 [2.1755–2.1962]	5.9554 [5.9455–5.9654]	0.367	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	none	device	poisson-32	cuda-native	11.1798 [11.0599–11.2998]	32.2885 [31.3239–33.2530]	0.346	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	none	device	poisson-64	cuda-native	26.1224 [25.6408–26.6040]	71.5496 [71.0574–72.0418]	0.365	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	none	device	poisson-128	cuda-native	35.7485 [34.1550–37.3421]	94.5589 [93.4998–95.6179]	0.378	395/404	395/404	5.661e-06	5.661e-06	5.749e-06
cuda-csr-f64-v1	jacobi	host	poisson-8	cuda-native	5.2566 [4.8006–5.7126]	10.1798 [8.9234–11.4361]	0.516	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	jacobi	host	poisson-32	cuda-native	27.5662 [25.2630–29.8694]	50.7118 [45.5096–55.9141]	0.544	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	jacobi	host	poisson-64	cuda-native	59.4428 [57.5221–61.3636]	109.9064 [108.4558–111.3570]	0.541	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	jacobi	host	poisson-128	cuda-native	80.1842 [78.7656–81.6027]	142.7977 [138.8050–146.7904]	0.562	395/404	395/404	5.661e-06	5.661e-06	5.749e-06
cuda-csr-f64-v1	jacobi	device	poisson-8	cuda-native	2.7074 [2.6891–2.7256]	9.0181 [8.5652–9.4709]	0.300	26/28	26/28	2.012e-07	2.012e-07	3.055e-07
cuda-csr-f64-v1	jacobi	device	poisson-32	cuda-native	12.9054 [12.6392–13.1717]	44.7771 [44.3602–45.1941]	0.288	129/133	129/133	1.328e-06	1.328e-06	1.367e-06
cuda-csr-f64-v1	jacobi	device	poisson-64	cuda-native	31.0486 [30.3585–31.7388]	105.9231 [99.5819–112.2643]	0.293	290/297	290/297	2.773e-06	2.773e-06	2.810e-06
cuda-csr-f64-v1	jacobi	device	poisson-128	cuda-native	42.0661 [41.0829–43.0493]	134.6804 [132.4652–136.8956]	0.312	395/404	395/404	5.661e-06	5.661e-06	5.749e-06

Every remaining native median at or above its same-artifact Torch reference¶
configuration	case	candidate	native ms	Torch ms	higher native median %
jacobi/api/fused=true/host	dense-64	cuda-cublas	3.0670	2.5677	19.45
jacobi/api/fused=true/host	dense-256	cuda-cublas	2.9675	2.0143	47.32
jacobi/api/fused=true/host	dense-1024	cuda-cublas	3.1502	2.2908	37.51
jacobi/api/fused=true/host	dense-2048	cuda-cublas	3.0547	2.1715	40.67
jacobi/api/fused=true/device	dense-64	cuda-cublas	2.6093	2.3936	9.01
jacobi/api/fused=true/device	dense-256	cuda-cublas	2.6720	2.0369	31.18
jacobi/api/fused=true/device	dense-1024	cuda-cublas	2.7173	2.2537	20.57
jacobi/api/fused=true/device	dense-2048	cuda-cublas	2.9176	2.2230	31.25

All 8 remaining losses are retained. Within-configuration work gate: true. Fresh-pass work changes: 0; cross-scope/mode work changes: 0. Differences in iterations/matvecs permit a time-to-accepted-tolerance interpretation, not a pure-kernel speed comparison.

Lifecycle and launch traces are separate diagnostics¶

All profiles are instrumented single-process completed-wall samples, with three warmups and ten captured repetitions. Their before/after independent true-residual gates pass. Counts below are totals across ten repetitions; resource-only probe work counts describe gated solves outside that probe. They are not framework scores, uninstrumented speed estimates, additive phase timings or fractions of end-to-end latency. CUDA API durations can include waiting for earlier GPU work and pageable-memory handling.

Workspace captures include construction and destruction; prepare captures include both the operator snapshot and Jacobi setup. Session construction is outside timed solves. Provider-internal allocations/synchronization remain observed costs. The old pinned host scalar allocation was already absent from this campaign’s baseline.

Baseline lifecycle captures¶
capture	compiled source	instrumented median ms	malloc/free	memcpy/stream-sync/device-sync	kernel launch API calls	gated iterations/matvecs	true residual	threshold
cuda-cublas-dense-64-jacobi-api-device-solve	6bfb7b4905e7	5.198334	90/100	120/140/30	770	5/7	1.141e-07	2.738e-06
cuda-cublas-dense-256-jacobi-api-device-solve	6bfb7b4905e7	3.917138	90/100	110/130/30	640	4/6	1.099e-06	6.342e-06
cuda-cublas-dense-1024-jacobi-api-device-solve	6bfb7b4905e7	4.496294	90/100	110/130/30	640	4/6	1.722e-07	1.079e-05
cuda-cublas-dense-2048-jacobi-api-device-solve	6bfb7b4905e7	4.661090	90/100	100/120/30	510	3/5	1.245e-05	1.474e-05
cuda-cublas-dense-1024-jacobi-api-host-solve	6bfb7b4905e7	4.115286	80/90	180/200/30	400	4/6	1.722e-07	1.079e-05
cuda-cublas-dense-1024-jacobi-api-host-workspace	6bfb7b4905e7	1.720404	50/60	0/20/30	0	4/6	1.722e-07	1.079e-05
cuda-cublas-dense-1024-jacobi-api-device-workspace	6bfb7b4905e7	1.752028	60/70	0/20/30	0	4/6	1.722e-07	1.079e-05
cuda-native-dense-1024-jacobi-api-device-solve	6bfb7b4905e7	1.568563	60/60	110/130/0	580	4/6	1.722e-07	1.079e-05
torch-cuda-eager-dense-1024-jacobi-api-device-solve	6bfb7b4905e7	3.193666	0/0	210/190/20	720	4/6	1.722e-07	1.079e-05
cuda-cublas-dense-1024-jacobi-api-device-prepare	6bfb7b4905e7	0.929155	30/30	10/40/0	10	4/6	1.722e-07	1.079e-05

Baseline lifecycle captures retained identities¶
capture	JSON SHA256	CUDA API CSV SHA256
cuda-cublas-dense-64-jacobi-api-device-solve	9539f4e5f54c38a39c77783d5c148edf7ff120ccf41818d5317cdacfd08b7a59	b7891b3fa67808f1b657291f550eae347e8dca1c2129e8b04ba5f8f39b71f9a0
cuda-cublas-dense-256-jacobi-api-device-solve	b3df9a92dd67aafe30b19f8e4a23a4125398980ed00f2d8ce34e7544d99d5881	f19e36007139b58441d50574e6701defd7c44537ccaf40a6040e2180fe56d5ba
cuda-cublas-dense-1024-jacobi-api-device-solve	6229dd3527f944ce8a7c0737e14a0431797c54f314ebb71afc757149a8d3b853	89cb42497fbf931dfcd21db4cdcb200482afd29f0de12ea4515c2bd5d0df1620
cuda-cublas-dense-2048-jacobi-api-device-solve	7c547b2f8fd1274a286bb2d61e00140889840c419dba2fbd9f7dfff4a866d2bb	f9fadf60c3f00232b43ef1657232b8efdce85cc7ac90bb7ac34adc720c3f6b7a
cuda-cublas-dense-1024-jacobi-api-host-solve	487f3e53891556b63801f75c26d3afa73b6acd1a419835f3635ff49e9974f71f	a6416c42e529f7d526ff014d82eeebc04c561d9dc8fe9e41a03caea614896956
cuda-cublas-dense-1024-jacobi-api-host-workspace	0ed8ffff099cf9896e413eba0ff85e33cf0ad79592346f32137d396f4d419775	0e7d0112c3b179a23a267df0751bc6be824979502f515b323022dafdde13f148
cuda-cublas-dense-1024-jacobi-api-device-workspace	7bdfc50ffd9e21df4e33fec521cd650e48137bd611611b615a9c197a866955e3	9d479726a424d67dc9298a3512df84d3ba7511b5bd700c3d2b47ae1c9beb2fca
cuda-native-dense-1024-jacobi-api-device-solve	d605644afdcb5bbde0e39ba80b0352a439fb0971ac1acc6bc2dcce0d8025549d	e450a062ef597c8e64fa673b5d620f5b7d30d791f81fed04fee66ae05b63965a
torch-cuda-eager-dense-1024-jacobi-api-device-solve	e303c6d6c25faee39cfc8d7244070d8596f4b6f4450ef7f6862bd8a778746204	4d56979de51d8b4170cc1b532036f43f54d8d6101b5aedbbf00e66a3e18a782d
cuda-cublas-dense-1024-jacobi-api-device-prepare	48d0297b228671ff5076671aada6efeceb7f5958da81548549d8bb65e1d58ccd	6f647bb485105e8d6ad1f4095fcf9639a560cf9a4483afb9857dee0c90d6754c

Resource/session/launch follow-up captures¶
capture	compiled source	instrumented median ms	malloc/free	memcpy/stream-sync/device-sync	kernel launch API calls	gated iterations/matvecs	true residual	threshold
session-cuda-cublas-dense-1024-api-host-jacobi-standard	03379431a8cc	3.471923	70/80	180/200/30	400	4/6	1.722e-07	1.079e-05
session-cuda-cublas-dense-1024-api-device-jacobi-standard	03379431a8cc	3.068955	80/90	110/130/30	640	4/6	1.722e-07	1.079e-05
session-cuda-native-dense-1024-session-device-jacobi-standard	03379431a8cc	1.294158	20/20	110/130/0	580	4/6	1.722e-07	1.079e-05
session-cuda-cublas-dense-1024-session-device-jacobi-standard	03379431a8cc	1.826763	20/20	110/130/0	640	4/6	1.722e-07	1.079e-05
session-torch-cuda-eager-dense-1024-session-device-jacobi-standard	03379431a8cc	2.411116	0/0	210/190/20	720	4/6	1.722e-07	1.079e-05
session-cuda-native-dense-1024-prepared-device-jacobi-standard	03379431a8cc	1.176872	0/0	100/120/0	570	4/6	1.722e-07	1.079e-05
session-cuda-cublas-dense-1024-prepared-device-jacobi-standard	03379431a8cc	1.088162	0/0	100/120/0	630	4/6	1.722e-07	1.079e-05
session-cuda-native-poisson-32-prepared-device-jacobi-standard	03379431a8cc	20.610954	0/0	1430/1410/0	15670	129/133	1.328e-06	1.367e-06
session-cuda-native-dense-1024-prepared-device-none-explicit	03379431a8cc	23.009121	0/0	1410/1440/0	10410	69/139	8.845e-06	1.079e-05
launch-cuda-cublas-dense-1024-api-device-jacobi-standard	ca0d6afebeda	2.909371	80/90	110/130/30	560	4/6	1.722e-07	1.079e-05
launch-cuda-cublas-dense-1024-session-device-jacobi-standard	ca0d6afebeda	1.038850	20/20	110/130/0	560	4/6	1.722e-07	1.079e-05
launch-cuda-cublas-dense-1024-prepared-device-jacobi-standard	ca0d6afebeda	1.147453	0/0	100/120/0	550	4/6	1.722e-07	1.079e-05
launch-cuda-native-dense-1024-prepared-device-jacobi-standard	ca0d6afebeda	1.329903	0/0	100/120/0	490	4/6	1.722e-07	1.079e-05
launch-cuda-native-poisson-32-prepared-device-jacobi-standard	ca0d6afebeda	17.301841	0/0	1430/1410/0	13090	129/133	1.328e-06	1.367e-06
launch-cuda-native-dense-1024-prepared-device-none-explicit	ca0d6afebeda	21.359759	0/0	1410/1440/0	9030	69/139	8.845e-06	1.079e-05

Resource/session/launch follow-up captures retained identities¶
capture	JSON SHA256	CUDA API CSV SHA256
session-cuda-cublas-dense-1024-api-host-jacobi-standard	e0c5ed29fc37c418b0212ad1d097c584ed80053cac87175657cc394d4a1d3e76	3061d6c3ea95f9672d0339fe8660fbeae0c7d1a95557442b0ba2f3d8dbd2891d
session-cuda-cublas-dense-1024-api-device-jacobi-standard	6e4d42ba06463ac96490dae5712150592af1927e8dfcfad8efa83ed69b9a7415	6cb2a0a20856484a1c1441641fa644ec6fef093761b8d32e9692c739ba8ed162
session-cuda-native-dense-1024-session-device-jacobi-standard	9397b5c74325551495e677ab35cb29ea8c3bf8533d10134c46d0e09232636311	f6abf7df3b8a8bf96757903a578f7eff0cc08f2b10d133bb99e956c3748f4400
session-cuda-cublas-dense-1024-session-device-jacobi-standard	009b0f835db32de37eb92061974d6c2622e25459fe62250b0709e3c1c4f8aad0	a5924d352c5b1c8679281bbaf00b8ac08926a1964b56d475da5a217aa77a6928
session-torch-cuda-eager-dense-1024-session-device-jacobi-standard	8e502d3766a39aae5c0a3b5b01d2e7828122b844f857b02505d5b6aa2b7baf70	d8b3f8ceb6f4dbe36ebd2a50716f1f41be12c0be2abb704d9f610beb9a3802cb
session-cuda-native-dense-1024-prepared-device-jacobi-standard	d09896749c1ceae71da77ef868c5f05f6cf070af2b786a3c7f13eef166f0d725	b42551c2a785da0977a7e8d98167d7f483d8fe0f9898acb6967149f0ae277561
session-cuda-cublas-dense-1024-prepared-device-jacobi-standard	f4f12182a9054384c3a00e98333f6da5c275035c7f21ecab24ecf005ad66b22b	9e301d3ce64301589cb091f4c9265b52d9770189682fe25d5ef2594d39b4445b
session-cuda-native-poisson-32-prepared-device-jacobi-standard	b1fa65627add53d22c0eedb42d785bffd6f1c610a2989f5c0e7dbab6771e4478	a44d161c5371031e15897d345fad1cd02cb0a19d5749527cad3940e135b8455d
session-cuda-native-dense-1024-prepared-device-none-explicit	7b0f7f501d4222e1be83b633f0d9d210d276a63eb651c35e4f6379ea0dbcb90a	29ae09af64d8d8eb138feb6b322114a11df85d6acab3f722252c23c492d3a753
launch-cuda-cublas-dense-1024-api-device-jacobi-standard	576229965be9b3c26d5ec18afdcbbd0930c50726ca7d475e8d9070ca2ec54a9e	f359fb47776084e439e64e094e225e633852def7992d761cf333ab84889d95ee
launch-cuda-cublas-dense-1024-session-device-jacobi-standard	a884433e519f36c791d4756b09ae67703d8d725c471a37a172def6696b850b1f	9867199c82440871395de7731041150a77a7781a9e90cdac60d8f4a48c3c2a44
launch-cuda-cublas-dense-1024-prepared-device-jacobi-standard	2a080b01bae7644fe27e7488b5af0fb7f87fd18972dc71a3db79f596fa30aa6f	cadbebb26e57a81c5515cbf22e1c8e5bca511d3876162c01f4e15b4fbcb9196d
launch-cuda-native-dense-1024-prepared-device-jacobi-standard	9c448be520c7d776fe28e7532d115fa36a362a39836812ad7295fbca17d57148	5d7b362669db514c44407c4fd2bcf7758e26e3f1ec744b255d590caec827a356
launch-cuda-native-poisson-32-prepared-device-jacobi-standard	e88eb741e1cd3b3d254a6ca3696005a47741bfad2f3da7e1e537fedf91324402	9bffa0e18134072ca9ad3ceaa5171c59b4bfc2f223ffafd4b5e1d02a54ee5fee
launch-cuda-native-dense-1024-prepared-device-none-explicit	7e967d05af3dcea160311c19ddd7304313594055a53892330ddcd6e9e9cdaa97	1c2e86eaa93f38e3e19a7d6224d425804319359427d5451d2cb24be27ff2f063

Compiled sources and retained evidence¶

All raw JSON, import sidecars, manifests, strict reports, frozen binaries and traces are local ignored evidence under build/execution-optimization-20261016/. They are not packaged or downloadable from these documentation links; a clean clone must rerun the public commands. Stage SHA256, actually loaded extension identities and input hashes are audited independently of the live collector checkout. Later rebases and CUDA lambda whitespace fixes do not replace the compiled identities below. No claim of identical whole trees or newly measured merge heads is made.

Frozen source, stage and active binary identities¶
stage identity	compiled revision	identity SHA256	active module	binary SHA256
build/execution-optimization-20261016/baseline-stages/cpu-release/identity.json	6bfb7b4905e79381a3d09772016aab7a860a0dad	17bc5a92fc367ab043af3aa0635156292b132d65e2beeeda67aface3d0bfb44b	_core	3111f94040afdea2519c5e61023336b8a7ef869c77bf84d7d6c4974eee8cbbff
build/execution-optimization-20261016/cpu-decoupled-stage/identity.json	7a7e6ea6e998a193514a1bb060d3cc67c0aaac45	4737c72f5f80f2cf87f4383a9bafebc162c493c90e93c2f0b327c7c7d019a9ab	_core	2e01054f7cbeff105546a5149374820b5a6a5ed1886840f750a8a431fcd3f678
build/execution-optimization-20261016/launch-stage/identity.json	ca0d6afebedaef1a1c4c7a5a2d2b4ffded6e1e2d	82ec5410a6d6763de6241b9002428d0e218f9b1261c74c99a6431d4c81fdc239	_core	2a6443b7d57a37eccb7a8436d8bd85618aadf4b867f3256025f4eafcefd6d2fa
build/execution-optimization-20261016/launch-stage/identity.json	ca0d6afebedaef1a1c4c7a5a2d2b4ffded6e1e2d	82ec5410a6d6763de6241b9002428d0e218f9b1261c74c99a6431d4c81fdc239	_cuda	754123ef7cf27270520b5c6a687a46c6b0645f7c5d595c8750efa9e3e363ddfd
build/execution-optimization-20261016/session-stage/identity.json	03379431a8cc4426447f1d5dfd011d908d9976bb	4fb7c4447f9502015a218f96c21fd5949b7d989a4c2a29916b36b2b043a6ac67	_core	2a6443b7d57a37eccb7a8436d8bd85618aadf4b867f3256025f4eafcefd6d2fa
build/execution-optimization-20261016/session-stage/identity.json	03379431a8cc4426447f1d5dfd011d908d9976bb	4fb7c4447f9502015a218f96c21fd5949b7d989a4c2a29916b36b2b043a6ac67	_cuda	819cb2911a36bb5cef4055741e371d21eb3386124612c8686106cd891ba74e46

Campaign manifest identities¶
campaign	path	manifest SHA256	fresh artifacts	eligible candidate/case results
CPU original ABBA	build/execution-optimization-20261016/cpu-decoupled-abba/manifest.json	5d2eb61c469c5a08ca98197ddc10281be8bbedf6f603e4c6badb47495f802a5a	32	384
CPU fixed confirmation	build/execution-optimization-20261016/cpu-none-api-confirmation/manifest.json	1fb60ac0d72e13d79f278f06144477dad8d3a24c219f7c4a42ee6cd6016917a8	8	96
CUDA three-scope characterization	build/execution-optimization-20261016/cuda-final-scopes/manifest.json	e033557af1d083adc50d17b80b3d74679213c009f7e06656c81f876f21170bde	48	480
CUDA launch ABBA	build/execution-optimization-20261016/launch-abba/manifest.json	42d8de7a2c6d229969046748ad5301669834f6473709dd292a0aa6dcc2f317dc	16	160

Baseline lifecycle captures manifest: build/execution-optimization-20261016/profiles-baseline/manifest.json, SHA256 130cebc1bc7e53427408638054fbefb5fb955c92e998321ee4538ac596d73e5d.

Baseline lifecycle captures profiler program and campaign runner are different files¶
role	recorded path	SHA256
profile_driver	scripts/profile_solve.py	e76030f99362e75913ee25b5aa0e0a6de640d073848f30c8f39000c6ae6cff2b
runner	build/execution-optimization-20261016/profile_baseline.py	c0c327660a7f23ce8e3fa565c6b45ee5691b315eed589a579eb58c58d806a8b0

retained runner SHA checked now; not recorded in historical manifest.

Resource/session/launch follow-up captures manifest: build/execution-optimization-20261016/profiles-followup/manifest.json, SHA256 1fdfbdca06a06fbec113ba42c046f888ca224f4abb7ab099123ec390e2e2c04c.

Resource/session/launch follow-up captures profiler program and campaign runner are different files¶
role	recorded path	SHA256
profile_driver	scripts/profile_solve.py	a3015a157af11f2af630f937f34efc480cc3398b17c5c25e1b7b4059ae824641
runner	build/execution-optimization-20261016/profile_followup.py	32d958e1b6b5c779169934058a3151389328fe262fef62efd2abd30f5c430a62

Saved strict report identities; original reports remain intact¶
campaign	configuration/scope	path	SHA256	overall verdict
CPU original ABBA	cpu-release-cpu-dense-f64-v1-none-prepared	build/execution-optimization-20261016/cpu-decoupled-abba/reports/cpu-release-cpu-dense-f64-v1-none-prepared.json	b5b3a1d6e9c8ab43a7ba8ba63f9341735e2afed066fc11f999e847a99f7a348a	inconclusive
CPU original ABBA	cpu-release-cpu-dense-f64-v1-none-api	build/execution-optimization-20261016/cpu-decoupled-abba/reports/cpu-release-cpu-dense-f64-v1-none-api.json	ef3253f6d846b684398c2fba2c504e1a255fe9140b594143bb9149ac2a1a0de8	regression
CPU original ABBA	cpu-release-cpu-dense-f64-v1-jacobi-prepared	build/execution-optimization-20261016/cpu-decoupled-abba/reports/cpu-release-cpu-dense-f64-v1-jacobi-prepared.json	17e2af458308413a5ef9e16eff0fdf8329eca3b32db46e67616bfb9a601b9c2e	inconclusive
CPU original ABBA	cpu-release-cpu-dense-f64-v1-jacobi-api	build/execution-optimization-20261016/cpu-decoupled-abba/reports/cpu-release-cpu-dense-f64-v1-jacobi-api.json	d0f66c497d6a36aea092b2ef90b43c29441fce01df86ed16bf570410e5307b76	inconclusive
CPU original ABBA	cpu-release-cpu-dense-f64-v1-none-prepared-unfused	build/execution-optimization-20261016/cpu-decoupled-abba/reports/cpu-release-cpu-dense-f64-v1-none-prepared-unfused.json	dcdb9e26351c2897675238ddccf0d87c1f5134ccbe740adf5eb46a1aa6228a6b	inconclusive
CPU original ABBA	cpu-release-cpu-dense-f64-v1-none-api-unfused	build/execution-optimization-20261016/cpu-decoupled-abba/reports/cpu-release-cpu-dense-f64-v1-none-api-unfused.json	16f92929558f4ebb7576608a0e6b2f85f90af7130ebc76a1d7dc8d954c6e308b	inconclusive
CPU original ABBA	cpu-release-cpu-dense-f64-v1-jacobi-prepared-unfused	build/execution-optimization-20261016/cpu-decoupled-abba/reports/cpu-release-cpu-dense-f64-v1-jacobi-prepared-unfused.json	3c2737cbf661c29f8373ef782d8ddcf6e2bf741861399e4f9a6878cf9fee9b31	inconclusive
CPU original ABBA	cpu-release-cpu-dense-f64-v1-jacobi-api-unfused	build/execution-optimization-20261016/cpu-decoupled-abba/reports/cpu-release-cpu-dense-f64-v1-jacobi-api-unfused.json	4e3b64228c444636599223a82047d1589bb95f968c27a7d137c9b27980db5b97	inconclusive
CPU fixed confirmation	confirmation-only	build/execution-optimization-20261016/cpu-none-api-confirmation/reports/confirmation-only.json	a934452c6c7f135b636cbf4d8e62ed7b4d2f0df6714dc8aba19201effc309e83	inconclusive
CPU fixed confirmation	pooled-original-plus-confirmation	build/execution-optimization-20261016/cpu-none-api-confirmation/reports/pooled-original-plus-confirmation.json	79803ef9c3e094653f50048a37b533b346188300f2c51a322a5d503d63cf6cc2	inconclusive
CUDA launch ABBA	cuda-release-cuda-dense-f64-v1-none-prepared-device	build/execution-optimization-20261016/launch-abba/reports/cuda-release-cuda-dense-f64-v1-none-prepared-device.json	0c4c20004782a12bc334997bce25964a95c77a2c12054daf7adeb93eb2eb57dd	inconclusive
CUDA launch ABBA	cuda-release-cuda-dense-f64-v1-jacobi-prepared-device	build/execution-optimization-20261016/launch-abba/reports/cuda-release-cuda-dense-f64-v1-jacobi-prepared-device.json	1960041f64c31e2479cdbbfb4c163452af0779eaa27b4debe24adf2d02629ebc	inconclusive
CUDA launch ABBA	cuda-release-cuda-csr-f64-v1-none-prepared-device	build/execution-optimization-20261016/launch-abba/reports/cuda-release-cuda-csr-f64-v1-none-prepared-device.json	ce84ae52b38c6d8522c2f38e96a56a107a65251139658b30f34ead2e0dec9c55	inconclusive
CUDA launch ABBA	cuda-release-cuda-csr-f64-v1-jacobi-prepared-device	build/execution-optimization-20261016/launch-abba/reports/cuda-release-cuda-csr-f64-v1-jacobi-prepared-device.json	1211f44bbc33e4e33b1f16bf13234dd7aa77fdb52fd16b054fc6ec556c8bde39	inconclusive

Historical CPU comparator retained for the original artifact schema¶
campaign	frozen comparison source	source SHA256	original source path	matching frozen Python package
CPU fixed confirmation	build/execution-optimization-20261016/frozen-cpu-comparator/compare_performance.py	956d0fc65dc463bb2edfdc800da2dbcbfe3049bc4ae6996844fc0c0c804f6124	scripts/compare_performance.py	build/execution-optimization-20261016/cpu-decoupled-stage/python

The CPU confirmation uses the exact historical comparator source relocated to ignored storage with its matching frozen Python harness. The live comparator’s later Session support does not rewrite historical data or change this campaign’s 5% / 95% / 10,000-resample policy. Pooled original and new blocks were also checked together for exact hardware/compiler/runtime controls and complete one-thread library-pool inventories.

Public reproduction¶

Follow Build, test and document and Installation for project-local dependencies, separate CPU/CUDA environments, the CUDA 12.8 toolkit and supported host compiler. Build and validate each source before starting quiet measurements. Use separate source/build directories; do not run make against the frozen evidence stages. Public commands recreate the protocol on the current host, not the original binary hashes or observed medians. The local orchestration scripts are evidence tools and are not required public production APIs.

The exact compiled sources are retained by public tags. Fetch and verify them before creating detached checkouts (OUT must be a new local directory):

PROJECT=$PWD
OUT="$PROJECT/build/execution-reproduction"
mkdir -p "$OUT"
git fetch origin refs/tags/benchmarks/gemv-fusion-20261016:refs/tags/benchmarks/gemv-fusion-20261016
test "$(git rev-parse refs/tags/benchmarks/gemv-fusion-20261016)" = 7a7e6ea6e998a193514a1bb060d3cc67c0aaac45
git fetch origin refs/tags/benchmarks/cuda-session-20261016:refs/tags/benchmarks/cuda-session-20261016
test "$(git rev-parse refs/tags/benchmarks/cuda-session-20261016)" = 03379431a8cc4426447f1d5dfd011d908d9976bb
git fetch origin refs/tags/benchmarks/cuda-launch-20261016:refs/tags/benchmarks/cuda-launch-20261016
test "$(git rev-parse refs/tags/benchmarks/cuda-launch-20261016)" = ca0d6afebedaef1a1c4c7a5a2d2b4ffded6e1e2d
git worktree add --detach "$OUT/baseline-source" 6bfb7b4905e79381a3d09772016aab7a860a0dad
git worktree add --detach "$OUT/cpu-source" 7a7e6ea6e998a193514a1bb060d3cc67c0aaac45
git worktree add --detach "$OUT/session-source" 03379431a8cc4426447f1d5dfd011d908d9976bb
git worktree add --detach "$OUT/launch-source" ca0d6afebedaef1a1c4c7a5a2d2b4ffded6e1e2d

CPU A/B uses baseline versus cpu-source. For every none/jacobi, prepared/api and fused/no-fused combination below, invoke four fresh processes in A,B,B,A order and save separate JSON files. A/B selects the source and build directory only; the entire candidate list and options stay the same. Build both sources with the recorded GNU 16 release configuration. Example for one block (set SRC, BUILD, RUN for each invocation):

taskset -c 0 make -C "$SRC" pyprof PRESET=cpu-release \
    PYTHON="$PROJECT/.venv/bin/python" BUILD_DIR="$BUILD" \
    PYPROF_SUITE=cpu-dense-f64-v1 PYPROF_SCOPE=api PYPROF_THREADS=1 \
    PYPROF_ROUNDS=30 PYPROF_REPETITIONS=3 PYPROF_OUTPUT="$OUT/$RUN.json" \
    PYPROF_ARGS='--implementations native,numpy-out,torch-eager-out --preconditioner none --fused --warmups 3 --seed 20261015 --rtol 1e-8 --atol 1e-12 --maxiter 2000 --residual-refresh 50'

Repeat the configuration loop with --preconditioner jacobi, scope prepared, and --no-fused. The one fixed fused none/api confirmation adds A,B,B,A,B,A,A,B once; its pooled comparison includes all original two plus all four new blocks per side, with no case/candidate filtering.

CUDA characterization uses launch-source with all three scopes, both native execution modes, both preconditioners and dense/CSR suites: 24 configurations in forward order, then the exact reverse order, each in a fresh process. The default host mode is explicit host on the CLI and omits the optional device-engine field from historical specs. Builds use GNU 13 with CUDA 12.8, sm75 and no fast-math. For example:

taskset -c 0 make -C "$OUT/launch-source" pyprof PRESET=cuda-release \
    PYTHON="$PROJECT/.venv-cuda/bin/python" BUILD_DIR="$OUT/launch-build" \
    CUDA_ROOT="$PROJECT/.deps/cuda-12.8" \
    CUDA_HOST_COMPILER=/usr/bin/g++-13 CUDA_ARCHITECTURES=75 \
    PYPROF_SUITE=cuda-dense-f64-v1 PYPROF_SCOPE=session PYPROF_THREADS=1 \
    PYPROF_ROUNDS=30 PYPROF_REPETITIONS=3 \
    PYPROF_OUTPUT="$OUT/cuda-dense-jacobi-session-device-r1.json" \
    PYPROF_ARGS='--device cuda --cuda-device 0 --cuda-execution-mode device --implementations cuda-native,cuda-cublas,torch-cuda-eager --preconditioner jacobi --fused --warmups 3 --seed 20261015 --rtol 1e-8 --atol 1e-12 --maxiter 2000 --residual-refresh 50'

Set PYPROF_SCOPE=api|session|prepared, native mode host|device and preconditioner none|jacobi separately. For CSR use PYPROF_SUITE=cuda-csr-f64-v1, PYPROF_REPETITIONS=1 and the candidates cuda-native,torch-cuda-eager. Keep all cases, including losses. Pin taskset -c 0 and all BLAS/OpenMP thread variables to one; the public worker enforces its declared controls. Set CUBLAS_WORKSPACE_CONFIG=:4096:8 and leave CUDA_LAUNCH_BLOCKING unset.

Launch A/B uses session-source versus launch-source in A,B,B,A order for prepared/device only, both suites and preconditioners, with unchanged full candidate lists and timing boundaries. Compare equal-spec blocks using the public strict comparator (the process exits nonzero for an inconclusive or regression verdict; preserve the report):

PYTHONPATH="$SRC/python" "$PROJECT/.venv/bin/python" \
    "$SRC/scripts/compare_performance.py" \
    --baseline "$OUT/A1.json" "$OUT/A4.json" \
    --current "$OUT/B2.json" "$OUT/B3.json" \
    --max-regression 0.05 --bootstrap-samples 10000 --seed 20261015 \
    --output "$OUT/comparison.json"

Never pass different scopes or execution modes to that comparator. A raw replay requires the actual raw JSON, its matching compiled source/runtime and the same public CLI; these ignored local artifacts are not shipped in a clean clone. Nsight Systems is optional host tooling, not a pip dependency. Recreate diagnostic captures with scripts/profile_solve.py and the nsys flags in cuBLAS API lifecycle diagnostics — 2026-10-16; select the exact source, implementation, case, scope, mode and preconditioner from the capture table. Resource/prepare probes remain distinct from solve captures.