NVIDIA A10 · fastforks


device	NVIDIA A10
backend	cuda
memory	22 GB
hardware	sm_86
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/a10-22gb.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`	178	–	97.6	1.8x	1115 / 1236	642 / 798
`aanet-sceneflow`	148	–	105	1.4x	819 / 922	502 / 620
`bridgedepth-sf`	368	–	333	1.1x	2516 / 2792	1905 / 2210
`c-foundationstereo-s`	3089	–	1537	2.0x	6285 / 6590	3078 / 4388
`cas-gwcnet-gc-sceneflow`	380	–	345	1.1x	2232 / 2446	2237 / 2450
`cfnet-sceneflow`	461	–	432	1.1x	3406 / 4336	3636 / 4838
`coex-sceneflow`	62.2	–	27.3	2.3x	1102 / 1214	420 / 554
`crestereo-eth3d`	661	–	255	2.6x	786 / 946	617 / 648
`defom-stereo-vitl-sceneflow`	–	1211	689	1.8x	4333 / 4792	3124 / 3522
`dlnr-sceneflow`	346	–	224	1.5x	2608 / 3016	1080 / 1262
`fast-acvnet-generalization`	54.3	–	43.6	1.2x	662 / 756	738 / 856
`fast-acvnet-plus-sceneflow`	71.7	–	51.2	1.4x	659 / 756	748 / 976
`fast-foundationstereo`	332	–	170	1.9x	2083 / 2456	1117 / 1296
`foundationstereo-vit-l`	2456	–	1674	1.5x	3114 / 3416	2788 / 3040
`foundationstereo-vit-s`	2126	–	1369	1.6x	2466 / 2726	2140 / 2590
`great-igev-sceneflow`	982	–	477	2.1x	1375 / 1558	1195 / 1622
`gwcnet-gc-sceneflow`	297	–	244	1.2x	1801 / 2102	1765 / 1942
`hitnet-eth3d`	65.4	–	28.3	2.3x	857 / 1088	632 / 770
`igev-plusplus-sceneflow`	953	–	463	2.1x	1568 / 1736	1333 / 1756
`igev-stereo-sceneflow`	668	–	349	1.9x	1210 / 1432	1115 / 1312
`las2-h`	144	–	65.1	2.2x	1023 / 1066	888 / 980
`las2-s`	56.2	–	16.9	3.3x	921 / 966	196 / 280
`mocha-v1-sceneflow`	931	–	527	1.8x	1606 / 1788	1508 / 1648
`monster-plusplus-mix-all`	1533	–	792	1.9x	2557 / 2834	2355 / 2714
`monster-sceneflow`	1513	–	783	1.9x	2310 / 2634	2273 / 2594
`nmrf-sceneflow`	196	–	173	1.1x	1457 / 1630	969 / 1090
`psmnet-sceneflow`	276	–	218	1.3x	2043 / 2362	1752 / 2082
`raft-stereo-middlebury`	643	–	313	2.1x	1361 / 1570	730 / 758
`rt-monster-plusplus-mix-all`	112	–	105	1.1x	2565 / 2618	2646 / 3120
`s2m2-s`	284	–	137	2.1x	1706 / 1908	655 / 942
`s2m2-xl`	1350	–	822	1.6x	5202 / 6154	2065 / 2294
`selective-igev-sceneflow`	875	–	410	2.1x	1211 / 1448	1112 / 1430
`waft-stereo-b`	970	–	422	2.3x	5749 / 5810	1898 / 2052
`waft-stereo-s`	605	–	283	2.1x	3266 / 3434	1090 / 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`	69.5	60.1	1.2x	2610 / 2708	2592 / 2690
`da3-large-1.1`	33.5	21.1	1.6x	963 / 1070	948 / 1030
`da3metric-large`	22.3	18.6	1.2x	787 / 828	783 / 828
`da3mono-large`	22.0	18.5	1.2x	787 / 828	783 / 828
`da3nested-giant-large-1.1`	90.3	78.1	1.2x	3247 / 3354	3235 / 3338
`depth-anything-v2-b`	16.2	13.7	1.2x	421 / 474	350 / 396
`depth-anything-v2-l`	43.4	40.1	1.1x	998 / 1092	917 / 1034
`depth-anything-v2-s`	7.99	5.43	1.5x	220 / 266	153 / 208
`depth-pro`	375	359	1.0x	3278 / 3458	3606 / 3698
`marigold-depth-v1-1`	341	260	1.3x	2756 / 2920	2331 / 2480
`metric3d-vit-large`	175	166	1.1x	1424 / 1576	1455 / 1528
`midas-dpt-beit-large-512`	73.1	74.1	1.0x	932 / 1058	942 / 1036
`moge-2-vitl-normal`	120	95.2	1.3x	1583 / 1654	1419 / 1530
`moge-2-vits-normal`	56.7	35.6	1.6x	987 / 1132	805 / 868
`moge-3-vitl`	858	276	3.1x	2128 / 2424	1533 / 1684
`unidepth-v2-vitl14`	78.8	77.0	1.0x	1129 / 1212	1129 / 1212
`unik3d-vitl`	90.9	78.4	1.2x	1141 / 1212	1087 / 1126
`zoedepth-nk`	128	92.9	1.4x	1876 / 1964	1682 / 1862

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`	0.945	0.616	1.5x	127 / 164	136 / 188
`convnextv2_base.fcmae_ft_in22k_in1k`	10.1	3.79	2.7x	276 / 296	270 / 320
`deit_base_patch16_224.fb_in1k`	2.68	1.77	1.5x	179 / 206	252 / 308
`densenet121.tv_in1k`	9.25	1.91	4.9x	29 / 50	35 / 92
`dinov2_vitl14_lc`	5.94	6.24	1.0x	599 / 608	842 / 890
`efficientnet_b0.tv_in1k`	4.63	1.09	4.2x	24 / 34	30 / 78
`efficientnet_b4.tv_in1k`	8.97	2.55	3.5x	65 / 94	66 / 120
`efficientnet_v2_s.tv_in1k`	9.42	2.47	3.8x	57 / 86	65 / 128
`efficientvit_b2.r288_in1k`	8.93	1.40	6.4x	61 / 124	72 / 170
`eva02_large_patch14_448.mim_m38m_ft_in22k_in1k`	23.1	16.9	1.4x	624 / 644	1204 / 1346
`fastvit_sa24.apple_dist_in1k`	6.57	1.73	3.8x	57 / 88	63 / 128
`inception_v3.tv_in1k`	4.79	1.74	2.7x	97 / 124	102 / 168
`mobilenet_v2.tv2_in1k`	2.76	0.580	4.8x	19 / 30	26 / 52
`mobilenet_v3_large.tv2_in1k`	3.57	0.777	4.6x	22 / 50	30 / 74
`mobilenetv1_100.ra4_e3600_r224_in1k`	1.63	0.448	3.6x	20 / 30	27 / 54
`mobilenetv4_conv_medium.e500_r256_in1k`	3.70	0.900	4.1x	31 / 62	39 / 84
`regnety_160.tv2_in1k`	7.83	2.27	3.4x	184 / 230	181 / 276
`resnet152.tv2_in1k`	8.04	2.48	3.2x	129 / 160	134 / 206
`resnet18.tv_in1k`	1.40	0.609	2.3x	35 / 68	42 / 110
`shufflenet_v2_x1_0.tv_in1k`	4.36	1.22	3.6x	15 / 28	39 / 56
`squeezenet1_1.tv_in1k`	1.82	0.461	4.0x	7 / 28	8 / 12
`swin_base_patch4_window7_224.ms_in22k_ft_in1k`	11.7	2.80	4.2x	187 / 210	198 / 258
`vgg16.tv_in1k`	1.64	1.47	1.1x	287 / 308	285 / 348
`vit_base_patch16_224.augreg2_in21k_ft_in1k`	2.85	1.77	1.6x	179 / 206	252 / 308
`vit_so400m_patch14_siglip_378.webli_ft_in1k`	16.4	16.0	1.0x	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`	40.8	20.8	2.0x	410 / 452	355 / 436
`dinov2-l-linear-ade20k`	70.3	67.1	1.0x	1154 / 1224	1162 / 1244
`dinov2-s-linear-ade20k`	16.6	13.1	1.3x	615 / 656	618 / 656
`eomt-l-dinov2-ade20k`	39.5	38.3	1.0x	1488 / 1564	1488 / 1564
`eomt-l-dinov3-ade20k`	45.9	44.6	1.0x	1487 / 1564	1487 / 1564
`fcn-r101-d8-ade20k`	33.8	19.5	1.7x	414 / 472	288 / 374
`mask2former-swinl-ade20k`	94.5	67.4	1.4x	1131 / 1358	1131 / 1358
`oneformer-swinl-ade20k`	133	109	1.2x	2022 / 2454	2022 / 2434
`sapiens-seg-0.3b-goliath`	99.3	97.7	1.0x	1616 / 1742	1616 / 1742
`sapiens-seg-1b-goliath`	210	200	1.0x	3232 / 3362	3233 / 3364
`segformer-b0-ade20k`	9.78	3.79	2.6x	312 / 336	89 / 178
`segformer-b5-ade20k`	39.4	28.5	1.4x	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`	39.7	35.1	1.1x	341 / 360	341 / 380
`sam-vit-h`	182	failed	–	1448 / 1504	–
`sam-vit-l`	99.7	failed	–	797 / 842	–
`sam2.1-hiera-base-plus`	34.2	failed	–	376 / 404	–
`sam2.1-hiera-large`	65.2	failed	–	706 / 728	–
`sam2.1-hiera-tiny`	17.7	15.6	1.1x	268 / 302	280 / 302
`sam3`	151	136	1.1x	1886 / 1964	1927 / 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`	18.0	11.6	1.6x	292 / 350	229 / 270
`detr-r50-coco`	13.1	8.64	1.5x	256 / 312	193 / 232
`dfine-l-coco`	15.7	9.48	1.7x	128 / 152	134 / 254
`dfine-n-coco`	10.2	3.68	2.8x	65 / 76	78 / 160
`dfine-x-coco`	19.7	14.4	1.4x	203 / 238	205 / 318
`faster-rcnn-r50-fpn-3x-coco`	22.8	18.1	1.3x	318 / 360	318 / 360
`mask-rcnn-r50-fpn-3x-coco`	24.5	20.1	1.2x	328 / 380	328 / 380
`rtdetrv2-l-coco`	12.2	7.63	1.6x	153 / 180	154 / 266
`rtdetrv2-s-coco`	8.33	4.88	1.7x	110 / 126	112 / 208
`rtdetrv2-x-coco`	15.0	10.4	1.4x	220 / 238	222 / 340

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`	19.5	18.0	1.1x	2191 / 2208	2198 / 2250
`dinov2-vitl14`	5.34	5.45	1.0x	598 / 604	606 / 668
`dinov2-vits14`	2.96	0.996	3.0x	54 / 58	63 / 82
`dinov3-convnext-large`	5.25	3.73	1.4x	395 / 410	393 / 432
`dinov3-convnext-tiny`	3.09	1.34	2.3x	68 / 88	59 / 90
`dinov3-vith16plus`	14.6	11.4	1.3x	1620 / 1628	1620 / 1672
`dinov3-vitl16`	9.35	3.76	2.5x	592 / 606	593 / 650
`dinov3-vits16`	4.71	1.02	4.6x	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`	488	483	1.0x	3623 / 3992	3623 / 3992
`vggt-1b`	517	504	1.0x	3619 / 3834	3630 / 4174
`vggt-omega-1b-512`	578	447	1.3x	3380 / 3476	3380 / 3434

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`	111	71.9	1.5x	847 / 948	786 / 840

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`	260	30.1	8.6x	1690 / 1892	624 / 752