NVIDIA L40S · fastforks


device	NVIDIA L40S
backend	cuda
memory	44 GB
hardware	sm_89
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	238 in `docs/data/speed/l40s-44gb.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`	80.7	–	46.0	1.8x	1115 / 1236	778 / 938
`aanet-sceneflow`	62.7	–	44.7	1.4x	819 / 922	502 / 620
`bridgedepth-sf`	135	–	122	1.1x	2516 / 2832	1905 / 2210
`c-foundationstereo-s`	1389	–	551	2.5x	6285 / 6570	3198 / 4328
`cas-gwcnet-gc-sceneflow`	193	–	159	1.2x	2232 / 2446	2255 / 2470
`cfnet-sceneflow`	175	–	152	1.2x	3406 / 4336	3628 / 4598
`coex-sceneflow`	21.9	–	11.9	1.8x	1102 / 1214	444 / 654
`crestereo-eth3d`	275	–	102	2.7x	786 / 946	617 / 648
`defom-stereo-vitl-sceneflow`	–	499	271	1.8x	4333 / 4812	3124 / 3542
`dlnr-sceneflow`	143	–	94.6	1.5x	2608 / 3016	1210 / 1382
`fast-acvnet-generalization`	25.3	–	19.2	1.3x	662 / 756	738 / 916
`fast-acvnet-plus-sceneflow`	31.3	–	20.3	1.5x	659 / 756	748 / 836
`fast-foundationstereo`	125	–	70.2	1.8x	2083 / 2456	1169 / 1496
`foundationstereo-vit-l`	968	–	605	1.6x	3114 / 3416	2791 / 3300
`foundationstereo-vit-s`	838	–	487	1.7x	2466 / 2726	2143 / 2612
`great-igev-sceneflow`	354	–	176	2.0x	1375 / 1558	1195 / 1582
`gwcnet-gc-sceneflow`	152	–	109	1.4x	1801 / 2102	1783 / 2322
`hitnet-eth3d`	31.2	–	11.0	2.8x	857 / 1088	632 / 770
`igev-plusplus-sceneflow`	337	–	177	1.9x	1568 / 1736	1333 / 1674
`igev-stereo-sceneflow`	237	–	134	1.8x	1210 / 1432	1115 / 1332
`las2-h`	50.0	–	36.3	1.4x	1023 / 1066	928 / 1000
`las2-s`	28.2	–	15.3	1.8x	921 / 966	255 / 340
`mocha-v1-sceneflow`	345	–	224	1.5x	1606 / 1768	1561 / 1788
`monster-plusplus-mix-all`	490	–	288	1.7x	2557 / 2834	2408 / 2654
`monster-sceneflow`	487	–	287	1.7x	2310 / 2634	2326 / 2634
`nmrf-sceneflow`	79.4	–	64.6	1.2x	1457 / 1630	969 / 1090
`psmnet-sceneflow`	132	–	90.6	1.5x	2043 / 2362	1752 / 2080
`raft-stereo-middlebury`	256	–	121	2.1x	1361 / 1570	730 / 758
`rt-monster-plusplus-mix-all`	48.3	–	50.1	1.0x	2565 / 2618	2646 / 3100
`s2m2-s`	125	–	56.8	2.2x	1706 / 1908	660 / 882
`s2m2-xl`	668	–	310	2.2x	5202 / 6154	2065 / 2294
`selective-igev-sceneflow`	327	–	163	2.0x	1211 / 1448	1112 / 1390
`waft-stereo-b`	456	–	146	3.1x	5749 / 5810	1866 / 2074
`waft-stereo-s`	268	–	100	2.7x	3266 / 3434	1100 / 1262

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`	30.8	22.5	1.4x	2610 / 2708	2592 / 2690
`da3-large-1.1`	17.5	8.49	2.1x	963 / 1070	948 / 1030
`da3metric-large`	8.02	7.78	1.0x	787 / 828	783 / 828
`da3mono-large`	7.80	7.66	1.0x	787 / 828	783 / 828
`da3nested-giant-large-1.1`	38.0	30.0	1.3x	3247 / 3354	3235 / 3338
`depth-anything-v2-b`	5.39	4.67	1.2x	421 / 474	350 / 396
`depth-anything-v2-l`	14.4	13.6	1.1x	998 / 1072	917 / 1034
`depth-anything-v2-s`	3.75	2.24	1.7x	220 / 268	153 / 208
`depth-pro`	151	133	1.1x	3278 / 3458	3606 / 3698
`marigold-depth-v1-1`	140	128	1.1x	2756 / 2920	2331 / 2480
`metric3d-vit-large`	63.8	56.9	1.1x	1424 / 1576	1455 / 1528
`midas-dpt-beit-large-512`	27.8	26.8	1.0x	932 / 1058	944 / 1016
`moge-2-vitl-normal`	44.8	35.2	1.3x	1583 / 1652	1419 / 1530
`moge-2-vits-normal`	24.9	18.0	1.4x	987 / 1132	805 / 868
`moge-3-vitl`	391	93.5	4.2x	2128 / 2444	1533 / 1684
`unidepth-v2-vitl14`	27.8	25.4	1.1x	1129 / 1212	1129 / 1212
`unik3d-vitl`	31.2	27.7	1.1x	1141 / 1192	1087 / 1126
`zoedepth-nk`	45.7	32.1	1.4x	1874 / 1964	1685 / 1860

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`	0.939	0.544	1.7x	127 / 164	136 / 188
`convnextv2_base.fcmae_ft_in22k_in1k`	11.6	2.64	4.4x	276 / 296	270 / 320
`deit_base_patch16_224.fb_in1k`	3.45	1.22	2.8x	179 / 206	252 / 308
`densenet121.tv_in1k`	11.5	1.40	8.3x	29 / 50	35 / 92
`dinov2_vitl14_lc`	6.51	2.76	2.4x	599 / 608	842 / 890
`efficientnet_b0.tv_in1k`	6.35	0.844	7.5x	25 / 34	30 / 78
`efficientnet_b4.tv_in1k`	11.4	1.70	6.7x	69 / 94	66 / 120
`efficientnet_v2_s.tv_in1k`	17.1	1.74	9.8x	57 / 86	65 / 128
`efficientvit_b2.r288_in1k`	13.3	1.01	13.1x	61 / 124	72 / 170
`eva02_large_patch14_448.mim_m38m_ft_in22k_in1k`	16.7	6.82	2.4x	624 / 644	1204 / 1346
`fastvit_sa24.apple_dist_in1k`	9.55	1.27	7.5x	57 / 88	63 / 128
`inception_v3.tv_in1k`	6.13	1.31	4.7x	97 / 124	102 / 168
`mobilenet_v2.tv2_in1k`	3.54	0.506	7.0x	19 / 30	26 / 52
`mobilenet_v3_large.tv2_in1k`	4.43	0.632	7.0x	23 / 50	30 / 74
`mobilenetv1_100.ra4_e3600_r224_in1k`	2.15	0.413	5.2x	20 / 30	27 / 54
`mobilenetv4_conv_medium.e500_r256_in1k`	4.81	0.722	6.7x	38 / 62	39 / 84
`regnety_160.tv2_in1k`	9.03	1.58	5.7x	184 / 230	181 / 276
`resnet152.tv2_in1k`	10.7	1.84	5.8x	129 / 160	134 / 206
`resnet18.tv_in1k`	1.65	0.519	3.2x	35 / 68	42 / 110
`shufflenet_v2_x1_0.tv_in1k`	4.26	0.874	4.9x	15 / 28	39 / 56
`squeezenet1_1.tv_in1k`	2.42	0.437	5.5x	7 / 28	8 / 12
`swin_base_patch4_window7_224.ms_in22k_ft_in1k`	12.6	2.02	6.2x	187 / 210	198 / 258
`vgg16.tv_in1k`	1.69	1.03	1.6x	287 / 308	285 / 348
`vit_base_patch16_224.augreg2_in21k_ft_in1k`	3.51	1.22	2.9x	179 / 206	252 / 308
`vit_so400m_patch14_siglip_378.webli_ft_in1k`	8.01	6.99	1.1x	866 / 904	1190 / 1286

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`	14.8	7.55	2.0x	410 / 452	355 / 436
`dinov2-l-linear-ade20k`	26.7	22.8	1.2x	1154 / 1224	1162 / 1244
`dinov2-s-linear-ade20k`	9.32	6.60	1.4x	615 / 656	618 / 656
`eomt-l-dinov2-ade20k`	16.3	13.9	1.2x	1488 / 1564	1488 / 1564
`eomt-l-dinov3-ade20k`	18.1	15.4	1.2x	1487 / 1564	1487 / 1564
`fcn-r101-d8-ade20k`	12.6	6.92	1.8x	414 / 472	288 / 374
`mask2former-swinl-ade20k`	48.4	32.8	1.5x	1131 / 1358	1131 / 1358
`oneformer-swinl-ade20k`	61.9	47.2	1.3x	2022 / 2454	2022 / 2434
`sapiens-seg-0.3b-goliath`	40.9	39.7	1.0x	1754 / 1882	1766 / 1882
`sapiens-seg-1b-goliath`	88.1	81.4	1.1x	3370 / 3502	3382 / 3522
`segformer-b0-ade20k`	7.25	1.76	4.1x	312 / 336	89 / 178
`segformer-b5-ade20k`	28.6	9.82	2.9x	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`	17.6	15.7	1.1x	341 / 360	341 / 380
`sam-vit-h`	68.4	failed	–	1448 / 1504	–
`sam-vit-l`	38.6	failed	–	797 / 842	–
`sam2.1-hiera-base-plus`	13.6	failed	–	376 / 404	–
`sam2.1-hiera-large`	25.0	failed	–	706 / 728	–
`sam2.1-hiera-tiny`	8.58	8.14	1.1x	268 / 302	280 / 302
`sam3`	72.2	48.6	1.5x	1886 / 1964	1926 / 2144

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`	13.5	9.91	1.4x	292 / 350	231 / 270
`detr-r50-coco`	10.1	8.21	1.2x	256 / 312	195 / 232
`dfine-l-coco`	19.4	5.16	3.8x	128 / 152	135 / 234
`dfine-n-coco`	11.8	2.41	4.9x	65 / 76	77 / 160
`dfine-x-coco`	21.6	6.81	3.2x	203 / 238	207 / 318
`faster-rcnn-r50-fpn-3x-coco`	11.3	10.6	1.1x	318 / 360	320 / 380
`mask-rcnn-r50-fpn-3x-coco`	12.6	11.8	1.1x	328 / 380	330 / 380
`rtdetrv2-l-coco`	14.3	3.82	3.8x	153 / 180	153 / 264
`rtdetrv2-s-coco`	9.62	2.76	3.5x	110 / 126	112 / 208
`rtdetrv2-x-coco`	17.9	4.89	3.7x	220 / 238	224 / 380

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`	11.4	6.66	1.7x	2191 / 2208	2198 / 2250
`dinov2-vitl14`	6.49	2.39	2.7x	598 / 604	606 / 668
`dinov2-vits14`	3.39	0.721	4.7x	54 / 58	63 / 82
`dinov3-convnext-large`	6.41	2.36	2.7x	395 / 410	393 / 432
`dinov3-convnext-tiny`	3.47	0.965	3.6x	68 / 88	59 / 90
`dinov3-vith16plus`	15.5	4.46	3.5x	1620 / 1628	1620 / 1672
`dinov3-vitl16`	11.1	2.27	4.9x	592 / 606	593 / 650
`dinov3-vits16`	5.69	0.727	7.8x	52 / 58	53 / 84

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`	185	179	1.0x	3623 / 3992	3623 / 3992
`vggt-1b`	196	189	1.0x	3619 / 3834	3640 / 4134
`vggt-omega-1b-512`	199	166	1.2x	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`	59.2	29.3	2.0x	847 / 968	996 / 1060

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`	82.5	11.5	7.2x	1690 / 1892	624 / 752