NVIDIA RTX PRO 6000 · fastforks


device	NVIDIA RTX PRO 6000
backend	cuda
memory	96 GB GDDR7
hardware	Blackwell Server Edition, sm_120, 600 W
software	torch 2.14.1+cu130, cuda 13.0, driver 610.57.04
torch (rows)	2.14.1+cu130
measured	2026-11-08 to 2026-11-08
rows	348 in `docs/data/speed/rtx-pro-6000-96gb.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`	46.3	–	22.1	2.1x	1139 / 1276	778 / 978
`aanet-sceneflow`	36.8	–	20.9	1.8x	843 / 982	502 / 620
`bridgedepth-rvc-pretrain`	71.5	–	68.8	1.0x	2957 / 3204	2347 / 2642
`bridgedepth-sf`	71.6	–	68.8	1.0x	2906 / 3192	2295 / 2592
`c-fast-foundationstereo`	72.8	–	37.3	2.0x	2107 / 2456	1193 / 1536
`c-foundationstereo-s`	747	–	319	2.3x	6309 / 6450	3102 / 4248
`cas-gwcnet-c-sceneflow`	61.0	–	46.7	1.3x	1972 / 2448	1690 / 1890
`cas-gwcnet-gc-sceneflow`	93.6	–	78.8	1.2x	2303 / 2768	2237 / 2452
`cfnet-sceneflow`	89.5	–	76.4	1.2x	3406 / 4336	3628 / 4598
`coex-sceneflow`	14.9	–	6.43	2.3x	1102 / 1194	420 / 514
`crestereo-eth3d`	128	–	60.9	2.1x	810 / 966	640 / 708
`defom-stereo-vitl-sceneflow`	–	266	146	1.8x	4357 / 4972	3148 / 3542
`dlnr-middlebury`	76.4	–	48.6	1.6x	2632 / 3216	1234 / 1342
`dlnr-sceneflow`	76.4	–	48.6	1.6x	2632 / 3216	1234 / 1342
`fast-acvnet-generalization`	15.4	–	9.33	1.6x	662 / 736	709 / 794
`fast-acvnet-kitti12`	15.4	–	9.58	1.6x	662 / 736	709 / 794
`fast-acvnet-kitti15`	15.4	–	9.34	1.6x	662 / 736	709 / 794
`fast-acvnet-plus-generalization`	21.6	–	11.7	1.8x	659 / 736	748 / 976
`fast-acvnet-plus-kitti12`	21.6	–	11.5	1.9x	659 / 736	748 / 976
`fast-acvnet-plus-kitti15`	21.6	–	11.6	1.9x	659 / 736	748 / 976
`fast-acvnet-plus-sceneflow`	21.6	–	11.8	1.8x	659 / 736	748 / 976
`fast-acvnet-sceneflow`	15.4	–	9.55	1.6x	662 / 736	709 / 794
`fast-foundationstereo`	72.7	–	36.8	2.0x	2107 / 2456	1193 / 1536
`foundationstereo-vit-l`	496	–	337	1.5x	3138 / 3436	2815 / 3180
`foundationstereo-vit-s`	426	–	270	1.6x	2490 / 2786	2167 / 2610
`great-igev-eth3d`	191	–	97.2	2.0x	1399 / 1598	1189 / 1602
`great-igev-kitti`	191	–	97.4	2.0x	1399 / 1598	1189 / 1602
`great-igev-middlebury`	338	–	143	2.4x	1872 / 2078	1551 / 1702
`great-igev-sceneflow`	191	–	98.3	1.9x	1399 / 1598	1189 / 1602
`gwcnet-g-sceneflow`	102	–	50.5	2.0x	2046 / 2140	1764 / 1942
`gwcnet-gc-sceneflow`	104	–	56.7	1.8x	2047 / 2142	1765 / 1942
`hitnet-eth3d`	14.4	–	6.22	2.3x	857 / 1068	632 / 770
`hitnet-middlebury-d400`	40.1	–	14.3	2.8x	1709 / 2174	1216 / 1514
`igev-plusplus-eth3d`	185	–	104	1.8x	1591 / 1876	1327 / 1594
`igev-plusplus-kitti12`	185	–	104	1.8x	1591 / 1916	1327 / 1574
`igev-plusplus-kitti15`	185	–	104	1.8x	1591 / 1876	1327 / 1754
`igev-plusplus-middlebury`	185	–	104	1.8x	1591 / 1876	1327 / 1754
`igev-plusplus-sceneflow`	185	–	104	1.8x	1591 / 1876	1327 / 1574
`igev-stereo-eth3d`	128	–	72.6	1.8x	1234 / 1452	1110 / 1332
`igev-stereo-kitti12`	128	–	72.2	1.8x	1234 / 1452	1110 / 1512
`igev-stereo-kitti15`	128	–	72.3	1.8x	1234 / 1452	1110 / 1512
`igev-stereo-middlebury`	128	–	72.6	1.8x	1234 / 1452	1110 / 1512
`igev-stereo-sceneflow`	128	–	72.6	1.8x	1234 / 1452	1110 / 1512
`las2-h`	28.8	–	15.5	1.8x	1047 / 1186	951 / 1040
`las2-l`	22.1	–	8.96	2.5x	1002 / 1054	332 / 422
`las2-m`	17.0	–	6.91	2.5x	937 / 982	263 / 376
`las2-s`	14.5	–	4.99	2.9x	921 / 966	255 / 380
`mocha-v1-sceneflow`	189	–	116	1.6x	1630 / 1760	1585 / 1688
`mocha-v2-sceneflow`	185	–	115	1.6x	1629 / 1866	1585 / 1788
`monster-mix-all`	275	–	164	1.7x	2334 / 2614	2350 / 2574
`monster-plusplus-mix-all`	285	–	167	1.7x	2580 / 2734	2432 / 2794
`monster-sceneflow`	275	–	164	1.7x	2334 / 2614	2350 / 2574
`nmrf-sceneflow`	43.1	–	37.7	1.1x	1481 / 1630	993 / 1110
`psmnet-sceneflow`	93.1	–	48.9	1.9x	2043 / 2422	1752 / 2080
`raft-stereo-eth3d`	148	–	63.6	2.3x	1385 / 1490	730 / 758
`raft-stereo-middlebury`	148	–	63.6	2.3x	1385 / 1490	730 / 758
`raft-stereo-sceneflow`	148	–	63.6	2.3x	1385 / 1490	730 / 758
`rt-monster-plusplus-mix-all`	25.3	–	23.6	1.1x	2589 / 2658	2641 / 3080
`s2m2-l`	204	–	91.7	2.2x	3242 / 3816	1247 / 1510
`s2m2-m`	134	–	61.1	2.2x	2297 / 2826	874 / 1122
`s2m2-s`	71.2	–	29.0	2.5x	1730 / 2008	683 / 902
`s2m2-xl`	352	–	168	2.1x	5226 / 6094	2089 / 2334
`selective-igev-eth3d-finetune`	170	–	89.3	1.9x	1235 / 1448	1106 / 1370
`selective-igev-eth3d-train`	170	–	89.3	1.9x	1235 / 1448	1106 / 1370
`selective-igev-kitti12`	170	–	89.2	1.9x	1235 / 1448	1106 / 1368
`selective-igev-kitti15`	170	–	89.1	1.9x	1235 / 1448	1106 / 1370
`selective-igev-middlebury-finetune`	170	–	89.4	1.9x	1235 / 1448	1106 / 1368
`selective-igev-middlebury-train`	170	–	89.2	1.9x	1235 / 1488	1106 / 1368
`selective-igev-sceneflow`	170	–	89.2	1.9x	1235 / 1488	1106 / 1350
`waft-stereo-b`	257	–	85.7	3.0x	5773 / 5850	1987 / 2272
`waft-stereo-s`	159	–	60.1	2.6x	3290 / 3454	1130 / 1242

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`	19.6	14.2	1.4x	2633 / 2728	2592 / 2690
`da3-large-1.1`	11.2	6.59	1.7x	987 / 1090	948 / 1030
`da3metric-large`	6.29	6.04	1.0x	811 / 868	807 / 848
`da3mono-large`	5.94	5.81	1.0x	811 / 868	807 / 848
`da3nested-giant-large-1.1`	26.2	19.7	1.3x	3271 / 3374	3235 / 3338
`depth-anything-v2-b`	4.37	4.00	1.1x	445 / 514	350 / 396
`depth-anything-v2-l`	9.81	8.61	1.1x	1022 / 1112	917 / 1034
`depth-anything-v2-s`	2.23	2.02	1.1x	244 / 288	153 / 208
`depth-pro`	77.9	68.9	1.1x	3301 / 3438	3630 / 3718
`marigold-depth-v1-1`	99.7	74.8	1.3x	2780 / 2940	2355 / 2500
`metric3d-vit-large`	41.3	38.9	1.1x	1447 / 1598	1479 / 1628
`midas-dpt-beit-large-512`	17.8	17.8	1.0x	955 / 1078	970 / 1076
`moge-2-vitl-normal`	27.4	24.1	1.1x	1607 / 1674	1443 / 1550
`moge-2-vits-normal`	15.4	11.0	1.4x	1011 / 1132	829 / 908
`moge-3-vitl`	147	61.5	2.4x	1682 / 1964	1557 / 1744
`unidepth-v2-vitl14`	18.4	18.8	1.0x	1153 / 1232	1156 / 1232
`unik3d-vitl`	20.3	19.1	1.1x	1165 / 1232	1111 / 1166
`zoedepth-nk`	29.5	19.2	1.5x	1983 / 2064	1711 / 1900

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`	0.340	0.240	1.4x	151 / 184	184 / 228
`convnextv2_base.fcmae_ft_in22k_in1k`	4.03	2.83	1.4x	299 / 316	318 / 380
`deit_base_patch16_224.fb_in1k`	1.11	1.02	1.1x	203 / 226	300 / 368
`densenet121.tv_in1k`	3.64	1.31	2.8x	53 / 70	83 / 132
`dinov2_vitl14_lc`	2.22	2.31	1.0x	623 / 648	889 / 950
`efficientnet_b0.tv_in1k`	1.89	0.736	2.6x	48 / 54	78 / 118
`efficientnet_b4.tv_in1k`	3.67	1.61	2.3x	89 / 114	109 / 178
`efficientnet_v2_s.tv_in1k`	3.55	1.62	2.2x	81 / 126	113 / 188
`efficientvit_b2.r288_in1k`	5.37	0.926	5.8x	85 / 144	120 / 210
`eva02_large_patch14_448.mim_m38m_ft_in22k_in1k`	6.80	5.46	1.2x	651 / 664	1233 / 1406
`fastvit_sa24.apple_dist_in1k`	2.88	1.31	2.2x	81 / 128	111 / 188
`inception_v3.tv_in1k`	1.77	1.26	1.4x	120 / 164	149 / 228
`mobilenet_v2.tv2_in1k`	1.16	0.346	3.4x	43 / 50	74 / 92
`mobilenet_v3_large.tv2_in1k`	1.46	0.489	3.0x	46 / 70	78 / 114
`mobilenetv1_100.ra4_e3600_r224_in1k`	0.725	0.253	2.9x	44 / 70	75 / 114
`mobilenetv4_conv_medium.e500_r256_in1k`	1.46	0.592	2.5x	54 / 82	87 / 124
`regnety_160.tv2_in1k`	3.34	1.46	2.3x	208 / 250	228 / 316
`resnet152.tv2_in1k`	3.02	1.82	1.7x	153 / 200	182 / 266
`resnet18.tv_in1k`	0.550	0.370	1.5x	59 / 88	89 / 150
`resnet50.tv2_in1k`	1.20	0.666	1.8x	87 / 98	116 / 164
`shufflenet_v2_x1_0.tv_in1k`	1.37	0.811	1.7x	39 / 48	87 / 116
`squeezenet1_1.tv_in1k`	0.655	0.269	2.4x	7 / 28	8 / 12
`swin_base_patch4_window7_224.ms_in22k_ft_in1k`	4.06	1.91	2.1x	211 / 230	246 / 298
`vgg16.tv_in1k`	0.739	0.622	1.2x	311 / 328	331 / 388
`vit_base_patch16_224.augreg2_in21k_ft_in1k`	1.16	1.02	1.1x	203 / 226	300 / 368
`vit_so400m_patch14_siglip_378.webli_ft_in1k`	4.80	5.17	0.9x	877 / 904	1214 / 1346

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
`deeplabv3-r101-d8-ade20k`	8.61	6.03	1.4x	450 / 512	396 / 514
`deeplabv3plus-r101-d8-ade20k`	9.28	5.21	1.8x	410 / 472	355 / 456
`dinov2-b-linear-ade20k`	7.70	6.86	1.1x	762 / 864	768 / 884
`dinov2-l-linear-ade20k`	14.7	13.8	1.1x	1178 / 1244	1186 / 1264
`dinov2-s-linear-ade20k`	5.11	3.90	1.3x	639 / 676	642 / 676
`eomt-l-dinov2-ade20k`	9.87	8.51	1.2x	1512 / 1604	1512 / 1604
`eomt-l-dinov3-ade20k`	10.9	9.54	1.1x	1511 / 1584	1511 / 1584
`fcn-r101-d8-ade20k`	7.82	5.21	1.5x	414 / 472	288 / 374
`hrnet-w48-fcn-ade20k`	8.45	4.64	1.8x	421 / 450	265 / 344
`mask2former-swinl-ade20k`	24.9	16.9	1.5x	1155 / 1338	1155 / 1398
`oneformer-swinl-ade20k`	37.7	31.4	1.2x	2046 / 2474	2046 / 2454
`pspnet-r101-d8-ade20k`	12.7	10.3	1.2x	414 / 452	359 / 474
`sapiens-seg-0.3b-goliath`	24.4	25.7	1.0x	1916 / 2042	1916 / 2042
`sapiens-seg-1b-goliath`	45.5	46.1	1.0x	3532 / 3664	3532 / 3662
`segformer-b0-ade20k`	3.18	1.24	2.6x	336 / 356	137 / 218
`segformer-b5-ade20k`	10.4	8.22	1.3x	543 / 590	360 / 492

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`	11.2	10.3	1.1x	365 / 400	365 / 420
`sam-vit-h`	38.7	35.1	1.1x	1472 / 1524	1467 / 1544
`sam-vit-l`	21.6	19.6	1.1x	821 / 882	818 / 880
`sam2.1-hiera-base-plus`	7.78	7.85	1.0x	400 / 444	414 / 464
`sam2.1-hiera-large`	13.8	14.4	1.0x	730 / 768	748 / 768
`sam2.1-hiera-small`	5.43	5.58	1.0x	305 / 346	317 / 346
`sam2.1-hiera-tiny`	5.15	5.11	1.0x	292 / 322	304 / 342
`sam3`	33.4	29.5	1.1x	1910 / 1984	1974 / 2224

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`	4.76	3.52	1.4x	316 / 370	257 / 290
`detr-r50-coco`	3.32	3.14	1.1x	280 / 332	221 / 252
`dfine-l-coco`	7.21	4.23	1.7x	152 / 172	186 / 294
`dfine-m-coco`	5.64	3.30	1.7x	126 / 142	163 / 282
`dfine-n-coco`	4.38	1.89	2.3x	88 / 116	126 / 200
`dfine-s-coco`	4.74	2.37	2.0x	105 / 128	143 / 230
`dfine-x-coco`	7.23	6.05	1.2x	227 / 258	245 / 378
`faster-rcnn-r50-fpn-3x-coco`	6.42	5.86	1.1x	342 / 400	346 / 380
`mask-rcnn-r50-fpn-3x-coco`	6.82	6.55	1.0x	352 / 440	356 / 420
`rtdetrv2-l-coco`	5.07	3.31	1.5x	177 / 200	207 / 322
`rtdetrv2-m-coco`	4.37	2.68	1.6x	156 / 170	180 / 272
`rtdetrv2-ms-coco`	4.00	2.57	1.6x	166 / 182	196 / 302
`rtdetrv2-s-coco`	3.55	2.21	1.6x	134 / 146	158 / 248
`rtdetrv2-x-coco`	6.43	4.07	1.6x	243 / 278	270 / 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-vitb14`	1.15	0.985	1.2x	204 / 226	237 / 288
`dinov2-vitb14-reg`	1.11	0.987	1.1x	206 / 226	237 / 288
`dinov2-vitg14`	5.24	4.95	1.1x	2215 / 2228	2245 / 2290
`dinov2-vitg14-reg`	5.29	4.94	1.1x	2217 / 2228	2245 / 2290
`dinov2-vitl14`	2.13	2.00	1.1x	622 / 644	654 / 708
`dinov2-vitl14-reg`	2.19	2.00	1.1x	624 / 646	654 / 708
`dinov2-vits14`	1.07	0.645	1.7x	78 / 98	111 / 142
`dinov2-vits14-reg`	1.08	0.639	1.7x	79 / 98	111 / 142
`dinov3-convnext-base`	1.78	2.03	0.9x	208 / 232	174 / 214
`dinov3-convnext-large`	1.88	2.54	0.7x	419 / 430	441 / 492
`dinov3-convnext-small`	1.81	1.78	1.0x	133 / 148	100 / 152
`dinov3-convnext-tiny`	1.07	0.973	1.1x	92 / 108	59 / 110
`dinov3-vitb16`	1.93	0.865	2.2x	200 / 206	201 / 250
`dinov3-vith16plus`	5.27	3.56	1.5x	1644 / 1668	1644 / 1692
`dinov3-vitl16`	3.56	1.99	1.8x	616 / 646	617 / 670
`dinov3-vits16`	1.87	0.654	2.9x	76 / 78	77 / 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`	96.5	97.6	1.0x	3647 / 3992	3647 / 4030
`vggt-1b`	104	106	1.0x	3643 / 3874	3654 / 4194
`vggt-omega-1b-512`	110	87.6	1.3x	3404 / 3556	3404 / 3534

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`	23.6	12.5	1.9x	871 / 988	1134 / 1180

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`	39.6	7.95	5.0x	1714 / 1892	648 / 772