NVIDIA A100-SXM4-40GB


device	NVIDIA A100-SXM4-40GB
backend	cuda
memory	40 GB
hardware	sm_80
software	torch 2.14.1+cu130, cuda 13.0, driver 610.57.04
torch (rows)	2.14.1+cu130
measured	2026-11-07 to 2026-11-08
rows	234 in `docs/data/speed/a100-sxm4-40gb.jsonl`

A survey, not a regression floor: the models and dtypes measured here, as recorded. Cells are median ms; every column names its dtype. unsupported = the card does not offer that mode or dtype on this backend, failed = it ran and raised, OOM = out of memory, – = not tested; (golden fail) = it ran but its result failed the golden check on this device (never counted as fastest). Speed-up = kernels against the fastest original and the fastest portable row of the same model (any dtype); memory = the peak of that mode's fastest row, measured the way the column says (not comparable with another device's). One input per model, the task's standard one. Each model links to its model page; benchmarks has the method and the reference device.

Protocol: first call, 3 warm-up, peak reset, median of 3 rounds x median of 5.

Stereo

Stereo, Middlebury 2014 half, 994x1482

model	time (ms): portable fp16	portable bf16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`aanet-plus-sceneflow`	98.8	–	50.8	1.9x	1115 / 1236	778 / 1078
`aanet-sceneflow`	82.0	–	49.6	1.7x	819 / 922	502 / 620
`bridgedepth-rvc-pretrain`	156	–	143	1.1x	2567 / 2864	1957 / 2262
`bridgedepth-sf`	155	–	141	1.1x	2516 / 2832	1905 / 2212
`c-fast-foundationstereo`	152	–	85.1	1.8x	2647 / 3016	1169 / 1476
`c-foundationstereo-s`	1362	–	668	2.0x	5788 / 6350	3134 / 4088
`cas-gwcnet-c-sceneflow`	105	–	87.9	1.2x	1795 / 1948	1690 / 1888
`cas-gwcnet-gc-sceneflow`	153	–	136	1.1x	2232 / 2446	2237 / 2450
`cfnet-sceneflow`	156	–	145	1.1x	3406 / 4336	3636 / 4838
`coex-sceneflow`	32.6	–	16.3	2.0x	1102 / 1214	447 / 634
`crestereo-eth3d`	283	–	108	2.6x	786 / 946	617 / 648
`defom-stereo-vitl-sceneflow`	–	475	248	1.9x	4333 / 4792	3124 / 3522
`dlnr-sceneflow`	141	–	95.4	1.5x	2608 / 3016	1210 / 1382
`fast-acvnet-generalization`	31.7	–	21.5	1.5x	662 / 756	738 / 956
`fast-acvnet-plus-sceneflow`	48.3	–	28.7	1.7x	659 / 756	748 / 876
`fast-foundationstereo`	152	–	85.3	1.8x	2647 / 3016	1169 / 1476
`foundationstereo-vit-l`	959	–	680	1.4x	2874 / 3156	2911 / 3300
`foundationstereo-vit-s`	828	–	561	1.5x	2225 / 2486	2263 / 2830
`great-igev-sceneflow`	383	–	196	2.0x	1375 / 1558	1195 / 1582
`gwcnet-gc-sceneflow`	166	–	97.7	1.7x	1801 / 2102	1765 / 1942
`hitnet-eth3d`	32.2	–	16.9	1.9x	857 / 1088	632 / 770
`igev-plusplus-sceneflow`	384	–	186	2.1x	1568 / 1736	1333 / 1674
`igev-stereo-sceneflow`	259	–	137	1.9x	1210 / 1432	1115 / 1332
`las2-h`	65.8	–	38.3	1.7x	1023 / 1066	928 / 1040
`las2-s`	30.5	–	15.7	1.9x	921 / 966	255 / 360
`mocha-v1-sceneflow`	383	–	227	1.7x	1606 / 1768	1561 / 1690
`monster-plusplus-mix-all`	598	–	289	2.1x	2557 / 2834	2408 / 2654
`monster-sceneflow`	581	–	286	2.0x	2310 / 2634	2326 / 2634
`nmrf-sceneflow`	93.6	–	84.4	1.1x	1457 / 1630	969 / 1090
`psmnet-sceneflow`	153	–	91.3	1.7x	2043 / 2362	1752 / 2082
`raft-stereo-middlebury`	275	–	110	2.5x	1361 / 1570	730 / 758
`rt-monster-plusplus-mix-all`	49.4	–	54.3	0.9x	2565 / 2618	2646 / 3040
`s2m2-s`	124	–	62.4	2.0x	1706 / 1908	655 / 942
`s2m2-xl`	511	–	312	1.6x	5202 / 6154	2065 / 2294
`selective-igev-sceneflow`	343	–	168	2.0x	1211 / 1448	1112 / 1390
`waft-stereo-b`	439	–	173	2.5x	5749 / 5810	1762 / 1952
`waft-stereo-s`	298	–	125	2.4x	3266 / 3434	1062 / 1142

Depth

Depth, Middlebury 2014 half, 994x1482

model	time (ms): portable fp16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`da3-giant-1.1`	45.0	25.0	1.8x	2610 / 2708	2592 / 2690
`da3mono-large`	15.5	12.8	1.2x	787 / 828	783 / 828
`depth-anything-v2-l`	16.9	16.1	1.1x	998 / 1072	917 / 1034
`depth-anything-v2-s`	7.17	4.63	1.5x	220 / 268	153 / 208
`depth-pro`	128	119	1.1x	3278 / 3458	3606 / 3698
`marigold-depth-v1-1`	177	200	0.9x	2756 / 2920	2521 / 2622
`metric3d-vit-large`	68.5	61.5	1.1x	1424 / 1576	1455 / 1528
`midas-dpt-beit-large-512`	28.5	28.5	1.0x	932 / 1058	944 / 1036
`moge-2-vits-normal`	26.5	22.1	1.2x	987 / 1132	805 / 868
`moge-3-vitl`	461	114	4.0x	2128 / 2482	1533 / 1684
`unidepth-v2-vitl14`	29.0	26.6	1.1x	1129 / 1212	1129 / 1212
`unik3d-vitl`	35.8	29.8	1.2x	1141 / 1212	1087 / 1126
`zoedepth-nk`	61.3	41.2	1.5x	1874 / 1964	1686 / 1880

Note: kernels row measured at the last commit, with the published dpt_fused@415b13f0e87b.

Classification

Classification, synthetic, 224x224

model	time (ms): portable fp16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`alexnet.tv_in1k`	0.955	0.591	1.6x	128 / 164	136 / 188
`convnextv2_base.fcmae_ft_in22k_in1k`	12.2	5.03	2.4x	276 / 296	270 / 320
`deit_base_patch16_224.fb_in1k`	3.53	2.21	1.6x	179 / 206	252 / 308
`densenet121.tv_in1k`	12.4	2.69	4.6x	31 / 50	35 / 92
`dinov2_vitl14_lc`	6.71	4.71	1.4x	599 / 608	842 / 890
`efficientnet_b0.tv_in1k`	6.35	1.51	4.2x	24 / 34	30 / 78
`efficientnet_b4.tv_in1k`	12.0	3.30	3.6x	65 / 94	66 / 120
`efficientnet_v2_s.tv_in1k`	12.3	3.29	3.7x	57 / 86	65 / 128
`efficientvit_b2.r288_in1k`	12.2	1.89	6.4x	61 / 124	72 / 170
`eva02_large_patch14_448.mim_m38m_ft_in22k_in1k`	16.3	10.3	1.6x	624 / 644	1204 / 1346
`fastvit_sa24.apple_dist_in1k`	8.28	2.49	3.3x	57 / 88	63 / 128
`inception_v3.tv_in1k`	6.31	2.40	2.6x	97 / 124	102 / 168
`mobilenet_v2.tv2_in1k`	3.82	0.768	5.0x	19 / 30	26 / 52
`mobilenet_v3_large.tv2_in1k`	4.67	1.05	4.4x	22 / 52	30 / 74
`mobilenetv1_100.ra4_e3600_r224_in1k`	2.10	0.574	3.7x	20 / 30	27 / 54
`mobilenetv4_conv_medium.e500_r256_in1k`	4.99	1.19	4.2x	31 / 62	39 / 84
`regnety_160.tv2_in1k`	9.30	2.70	3.4x	184 / 230	181 / 276
`resnet152.tv2_in1k`	10.5	3.42	3.1x	129 / 160	134 / 206
`resnet18.tv_in1k`	1.71	0.797	2.1x	35 / 68	42 / 110
`shufflenet_v2_x1_0.tv_in1k`	4.41	1.63	2.7x	15 / 28	39 / 56
`squeezenet1_1.tv_in1k`	2.25	0.592	3.8x	7 / 28	8 / 12
`swin_base_patch4_window7_224.ms_in22k_ft_in1k`	12.8	3.79	3.4x	187 / 210	198 / 258
`vgg16.tv_in1k`	1.54	1.15	1.3x	287 / 308	285 / 348
`vit_base_patch16_224.augreg2_in21k_ft_in1k`	3.52	2.21	1.6x	179 / 206	252 / 308
`vit_so400m_patch14_siglip_378.webli_ft_in1k`	7.96	8.73	0.9x	853 / 864	1174 / 1266

Segmentation

Segmentation, synthetic, 1024x2048

model	time (ms): portable fp16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`deeplabv3plus-r101-d8-ade20k`	18.8	13.0	1.5x	410 / 452	355 / 436
`dinov2-l-linear-ade20k`	25.1	22.3	1.1x	1154 / 1224	1162 / 1244
`dinov2-s-linear-ade20k`	10.1	7.64	1.3x	615 / 656	618 / 656
`eomt-l-dinov2-ade20k`	17.0	15.0	1.1x	1488 / 1564	1488 / 1564
`eomt-l-dinov3-ade20k`	21.9	17.8	1.2x	1487 / 1564	1487 / 1564
`fcn-r101-d8-ade20k`	15.5	11.8	1.3x	414 / 472	288 / 374
`mask2former-swinl-ade20k`	52.1	35.3	1.5x	1131 / 1358	1131 / 1358
`oneformer-swinl-ade20k`	71.6	58.9	1.2x	2022 / 2454	2022 / 2434
`sapiens-seg-0.3b-goliath`	36.3	34.4	1.1x	1760 / 1882	1760 / 1882
`sapiens-seg-1b-goliath`	71.7	67.8	1.1x	3376 / 3502	3376 / 3502
`segformer-b0-ade20k`	8.68	3.43	2.5x	312 / 336	89 / 178
`segformer-b5-ade20k`	30.6	18.3	1.7x	520 / 570	337 / 452

Promptable segmentation

Promptable segmentation, synthetic, 1024x1024

model	time (ms): portable fp16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`sam-vit-b`	20.0	19.2	1.0x	341 / 360	341 / 380
`sam-vit-h`	67.4	61.3	1.1x	1448 / 1504	1443 / 1484
`sam-vit-l`	40.9	37.5	1.1x	797 / 842	795 / 862
`sam2.1-hiera-base-plus`	15.8	16.9	0.9x	376 / 404	390 / 424
`sam2.1-hiera-large`	27.1	28.0	1.0x	706 / 728	724 / 788
`sam2.1-hiera-tiny`	12.0	12.2	1.0x	268 / 302	280 / 302
`sam3`	69.7	50.8	1.4x	1886 / 1964	1926 / 2144

Note: kernels row measured at the last commit, with the published dpt_fused@415b13f0e87b.

Detection

Detection, synthetic, 800x1344

model	time (ms): portable fp16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`detr-r101-coco`	13.8	10.6	1.3x	292 / 350	229 / 270
`detr-r50-coco`	10.7	8.82	1.2x	256 / 312	193 / 232
`dfine-l-coco`	19.9	9.36	2.1x	128 / 152	135 / 234
`dfine-n-coco`	12.5	4.06	3.1x	65 / 76	81 / 160
`dfine-x-coco`	22.2	13.1	1.7x	203 / 238	206 / 338
`faster-rcnn-r50-fpn-3x-coco`	15.2	13.0	1.2x	318 / 360	318 / 360
`mask-rcnn-r50-fpn-3x-coco`	15.0	14.7	1.0x	328 / 380	328 / 380
`rtdetrv2-l-coco`	14.6	6.65	2.2x	153 / 180	157 / 262
`rtdetrv2-s-coco`	9.88	4.74	2.1x	110 / 126	113 / 208
`rtdetrv2-x-coco`	18.0	8.11	2.2x	220 / 238	224 / 360

Features

Features, synthetic, 224x224

model	time (ms): portable fp16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`dinov2-vitg14`	10.2	9.51	1.1x	2191 / 2208	2198 / 2250
`dinov2-vitl14`	7.04	3.77	1.9x	598 / 606	606 / 668
`dinov2-vits14`	3.33	1.40	2.4x	55 / 58	63 / 102
`dinov3-convnext-large`	6.18	4.10	1.5x	395 / 410	393 / 432
`dinov3-convnext-tiny`	3.53	1.79	2.0x	68 / 88	59 / 90
`dinov3-vith16plus`	15.4	6.54	2.3x	1620 / 1628	1620 / 1672
`dinov3-vitl16`	11.2	3.64	3.1x	593 / 606	593 / 650
`dinov3-vits16`	5.80	1.47	4.0x	53 / 58	53 / 104

Multiview

Multiview, synthetic, 336x518

model	time (ms): portable fp16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`mapanything-apache`	183	170	1.1x	3623 / 3992	3623 / 3992
`vggt-1b`	190	184	1.0x	3619 / 3834	3634 / 4134
`vggt-omega-1b-512`	210	150	1.4x	3380 / 3476	3380 / 3494

Multiview, synthetic, 448x672

model	time (ms): portable fp16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`mvsanywhere`	57.8	56.9	1.0x	847 / 948	1004 / 1140

Multiview, synthetic, 1536x1024

model	time (ms): portable fp16	kernels fp16	kernels speed-up: vs portable	peak memory (MiB, allocated / reserved, fastest dtype): portable	kernels
`patchmatchnet`	87.9	21.9	4.0x	1690 / 1892	624 / 752