U50/U50LV Performance - 1.3 English

Vitis AI Library User Guide (UG1354)
Document ID
UG1354
Release Date
2021-02-03
Version
1.3 English

The Xilinx® Alveo U50 Data Center accelerator cards are peripheral component interconnect express ( PCIe® ) Gen3x16 compliant and Gen4x8 compatible cards featuring the Xilinx 16 nm UltraScale+ technology. In this release, DPU is implemented in program logic for deep learning inference acceleration.

Note: Some models cannot run at the highest frequency of DPU and need DPU frequency reduction. See For Edge for DPU frequency reduction operation.

U50 Performance with 6E300 MHz DPUCAHX8H

Refer to the following table for the throughput performance (in frames/sec or fps) for various neural network samples on U50 Gen3x4 with DPUCAHX8H running at 6E@300 MHz.

Table 1. U50 Performance with 6E300 MHz DPUCAHX8H
No Neural Network Input Size GOPS DPU Frequency (MHz) Performance (fps) (Multiple thread)
1 densebox_320_320 320x320 0.49 300 2004.5
2 densebox_640_360 360x640 1.1 300 893.2
3 ENet_cityscapes_pt 512x1024 8.6 300x0.9 77.9
4 face_landmark 96x72 0.14 300 10042.2
5 face-quality 80x60 0.06 300x0.9 16522.1
6 face-quality_pt 80x60 0.06 300x0.9 16525.5
7 facerec_resnet20 112x96 3.5 300 1355.7
8 facerec-resnet20_mixed_pt 112x96 3.5 300x0.9 1235.2
9 facerec_resnet64 112x96 11 300 507.9
10 facereid-large_pt 96x96 0.5 300x0.9 7430.5
11 facereid-small_pt 80x80 0.09 300x0.9 18563
12 fpn 256x512 8.9 300 448.6
13 FPN_Res18_Medical_segmentation 320x320 45.3 300 103.5
14 FPN-resnet18_covid19-seg_pt 352x352 22.7 300x0.9 212.9
15 inception_resnet_v2_tf 299x299 26.4 300 186.1
16 inception_v1 224x224 3.2 300 1265.4
17 inception_v1_tf 224x224 3 300 1282.1
18 inception_v2 224x224 4 300 1004.8
19 inception_v3 299x299 11.4 300 419.1
20 inception_v3_pt 299x299 5.7 300 419
21 inception_v3_tf 299x299 11.5 300 418.8
22 inception_v3_tf2 299x299 11.5 300x0.9 377
23 inception_v4 299x299 24.5 300 194.2
24 inception_v4_2016_09_09_tf 299x299 24.6 300 194.3
25 medical_seg_cell_tf2 128x128 5.3 300x0.9 1063.9
26 MLPerf_resnet50_v1.5_tf 224x224 8.19 300x0.9 563.4
27 mlperf_ssd_resnet34_tf 1200x1200 433 300x0.9 14.1
28 multi_task 288x512 14.8 300 351.9
29 openpose_pruned_0_3 368x368 49.9 300x0.9 29
30 personreid-res18_pt 176x80 1.1 300x0.9 3542.2
31 personreid-res50_pt 256x128 5.4 300x0.9 915.8
32 plate_detection 320x320 0.49 300 5081.7
33 plate_num 96x288 1.75 300x0.9 1031.1
34 refinedet_baseline 480x360 123 300x0.9 50
35 RefineDet-Medical_EDD_tf 320x320 9.8 300x0.9 423.3
36 refinedet_pruned_0_8 360x480 25 300x0.9 198.3
37 refinedet_pruned_0_92 360x480 10.1 300x0.9 413.4
38 refinedet_pruned_0_96 360x480 5.1 300x0.9 608.3
39 refinedet_VOC_tf 320x320 81.9 300x0.9 70.9
40 reid 80x160 0.95 300 4000
41 resnet18 224x224 3.7 300 1480.9
42 resnet50 224x224 7.7 300 647
43 resnet50_pt 224x224 4.1 300 624.2
44 resnet50_tf2 224x224 7.7 300x0.9 583.8
45 resnet_v1_101_tf 224x224 14.4 300 367.8
46 resnet_v1_152_tf 224x224 21.8 300 246.1
47 resnet_v1_50_tf 224x224 7 300 712.4
48 salsanext_pt 64x2048 20.4 300x0.9 132.7
49 SemanticFPN_cityscapes_pt 256x512 10 300x0.9 454.5
50 semantic_seg_citys_tf2 512x1024 54 300x0.9 51.9
51 sp_net 128x224 0.55 300 1951.6
52 squeezenet 227x227 0.76 300 3538.5
53 squeezenet_pt 224x224 0.82 300 2235.9
54 ssd_adas_pruned_0_95 360x480 6.3 300 666.2
55 ssd_pedestrian_pruned_0_97 360x360 5.9 300x0.9 566.8
56 ssd_resnet_50_fpn_coco_tf 640x640 178.4 300x0.9 32.7
57 ssd_traffic_pruned_0_9 360x480 11.6 300 435.7
58 tiny_yolov3_vmss 416x416 5.46 300x0.9 872.3
59 unet_chaos-CT_pt 512x512 23.3 300x0.9 59.6
60 vgg_16_tf 224x224 31 300 161.7
61 vgg_19_tf 224x224 39.3 300 134.9
62 vpgnet_pruned_0_99 480x640 2.5 300x0.9 471.2
63 yolov2_voc 448x448 34 300x0.9 164
64 yolov2_voc_pruned_0_66 448x448 11.6 300x0.9 413.3
65 yolov2_voc_pruned_0_71 448x448 9.9 300x0.9 484.7
66 yolov2_voc_pruned_0_77 448x448 7.8 300x0.9 589.3
67 yolov3_adas_pruned_0_9 256x512 5.5 300x0.9 681.3
68 yolov3_bdd 288x512 53.7 300x0.9 78.3
69 yolov3_voc 416x416 65.4 300x0.9 80.7
70 yolov3_voc_tf 416x416 65.6 300x0.9 80.7
71 yolov4_leaky_spp_m 416x416 60.1 300x0.9 84.5

Model End-to-End Performance on U50 333 MHz Single Core DPUCAHX8L

Refer to the following table for the throughput performance (in frames/sec or fps) for various neural network samples on U50 Gen3x4 with DPUCAHX8L running at 1E@333 MHz.

Table 2. Model End-to-End Performance on U50 333 MHz Single Core DPUCAHX8L
No Neural Network Input Size GOPS Performance (fps) (Single thread) Performance (fps) (Multiple thread)
1 densebox_320_320 320x320 0.49 222.2 373.0
2 densebox_640_360 360x640 1.1 106.6 169.1
3 ENet_cityscapes_pt 512x1024 8.6 4.5 5.2
4 face_landmark 96x72 0.14 1955.1 5019.7
5 face-quality 80x60 0.06 2543.3 7674.0
6 face-quality_pt 80x60 0.06 2597.0 7729.3
7 facerec_resnet20 112x96 3.5 266.4 309.7
8 facerec-resnet20_mixed_pt 112x96 3.5 268.3 311.9
9 facerec_resnet64 112x96 11 132.2 145.6
10 facereid-small_pt 80x80 0.09 1773.0 4935.3
11 fpn 256x512 8.9 27.3 33.1
12 FPN_Res18_Medical_segmentation 320x320 45.3 13.7 14.5
13 FPN-resnet18_covid19-seg_pt 352x352 22.7 77.8 90.5
14 inception_resnet_v2_tf 299x299 26.4 31.8 33.7
15 inception_v1 224x224 3.2 248.0 330.3
16 inception_v1_tf 224x224 3 238.6 335.4
17 inception_v2 224x224 3.88 166.3 200.6
18 inception_v3 299x299 11.4 93.6 109.0
19 inception_v3_pt 299x299 5.7 92.2 111.3
20 inception_v3_tf 299x299 11.5 92.4 110.9
21 inception_v3_tf2 299x299 11.5 87.0 101.4
22 inception_v4 299x299 24.5 47.2 51.0
23 inception_v4_2016_09_09_tf 299x299 24.6 46.6 51.0
24 medical_seg_cell_tf2 128x128 5.3 75.0 76.6
25 MLPerf_resnet50_v1.5_tf 224x224 8.19 80.4 89.4
26 mlperf_ssd_resnet34_tf 1200x1200 433 4.4 7.0
27 mobilenet_1_0_224_tf2 224x224 1.1 649.5 2014.0
28 mobilenet_v1_0_5_160_tf 160x160 0.15 1147.7 5413.3
29 mobilenet_v1_1_0_224_tf 224x224 1.1 654.7 2085.3
30 mobilenet_v2 224x224 0.6 580.0 1156.2
31 mobilenet_v2_1_0_224_tf 224x224 0.6 528.0 1152.4
32 mobilenet_v2_1_4_224_tf 224x224 1.2 447.8 860.6
33 multi_task 288x512 14.8 14.4 17.9
34 openpose_pruned_0_3 368x368 49.9 10.4 17.2
35 personreid-res50_pt 256x128 5.4 97.6 107.3
36 plate_detection 320x320 0.49 462.5 1220.6
37 refinedet_baseline 480x360 123 27.4 30.7
38 RefineDet-Medical_EDD_tf 320x320 9.8 93.8 157.7
39 refinedet_pruned_0_8 360x480 25 56.9 71.2
40 refinedet_pruned_0_92 360x480 10.1 62.7 76.6
41 refinedet_pruned_0_96 360x480 5.1 72.4 94.8
42 refinedet_VOC_tf 320x320 81.9 30.8 48.2
43 reid 80x160 0.95 444.5 597.7
44 resnet18 224x224 3.7 229.0 290.4
45 resnet50 224x224 7.7 80.9 87.9
46 resnet50_pt 224x224 4.1 80.8 88.3
47 resnet50_tf2 224x224 7.7 81.0 87.2
48 resnet_v1_101_tf 224x224 14.4 53.3 55.5
49 resnet_v1_152_tf 224x224 21.8 37.1 38.6
50 resnet_v1_50_tf 224x224 7 92.3 102.8
51 retinaface 360x640 1.11 96.8 195.1
52 salsanext_pt 64x2048 20.4 7.2 10.5
53 SemanticFPN_cityscapes_pt 256x512 10 27.5 35.3
54 semantic_seg_citys_tf2 512x1024 54 4.6 5.4
55 sp_net 128x224 0.55 1476.5 7606.3
56 squeezenet 227x227 0.76 441.9 756.3
57 squeezenet_pt 224x224 0.82 311.0 466.6
58 ssd_adas_pruned_0_95 360x480 6.3 75.8 125.0
59 ssdlite_mobilenet_v2_coco_tf 300x300 1.5 273.8 609.8
60 ssd_mobilenet_v1_coco_tf 300x300 2.5 338.4 944.6
61 ssd_mobilenet_v2 360x480 6.6 41.7 158.6
62 ssd_mobilenet_v2_coco_tf 300x300 3.8 158.1 238.4
63 ssd_pedestrian_pruned_0_97 360x360 5.9 25.9 30.9
64 ssd_traffic_pruned_0_9 360x480 11.6 63.0 114.4
65 vgg_16_tf 224x224 31 53.6 57.6
66 vgg_19_tf 224x224 39.3 48.4 51.3
67 vpgnet_pruned_0_99 480x640 2.5 19.1 20.0

U50LV Performance with 9E275 MHz DPUCAHX8H

The following table shows the throughput performance (in frames/sec or fps) for various neural network samples on U50LV Gen3x4 with DPUCAHX8H running at 9E@275 MHz.

Table 3. U50LV Performance with 9E275 MHz DPUCAHX8H
No Neural Network Input Size GOPS DPU Frequency (MHz) Performance (fps) (Multiple thread)
1 densebox_320_320 320x320 0.49 275 2304.9
2 densebox_640_360 360x640 1.1 275 1025.6
3 ENet_cityscapes_pt 512x1024 8.6 275x0.9 85.5
4 face_landmark 96x72 0.14 275 13366.1
5 face-quality 80x60 0.06 275x0.9 18244.3
6 face-quality_pt 80x60 0.06 275x0.9 18103
7 facerec_resnet20 112x96 3.5 275 1831
8 facerec-resnet20_mixed_pt 112x96 3.5 275x0.9 1667.5
9 facerec_resnet64 112x96 11 275 676.7
10 facereid-large_pt 96x96 0.5 275x0.9 9524.1
11 facereid-small_pt 80x80 0.09 275x0.9 19880
12 fpn 256x512 8.9 275x0.9 566.5
13 FPN_Res18_Medical_segmentation 320x320 45.3 275 140.5
14 FPN-resnet18_covid19-seg_pt 352x352 22.7 275x0.9 287.4
15 inception_resnet_v2_tf 299x299 26.4 275 238.3
16 inception_v1 224x224 3.2 275 1730.8
17 inception_v1_tf 224x224 3 275 1749.8
18 inception_v2 224x224 4 275 1377
19 inception_v3 299x299 11.4 275 575.4
20 inception_v3_pt 299x299 5.7 275 575.4
21 inception_v3_tf 299x299 11.5 275 575.3
22 inception_v3_tf2 299x299 11.5 275x0.9 517.8
23 inception_v4 299x299 24.5 275 264.8
24 inception_v4_2016_09_09_tf 299x299 24.6 275 264.8
25 medical_seg_cell_tf2 128x128 5.3 275x0.9 1457.8
26 MLPerf_resnet50_v1.5_tf 224x224 8.19 275x0.9 689.4
27 mlperf_ssd_resnet34_tf 1200x1200 433 275x0.9 18.9
28 multi_task 288x512 14.8 275x0.9 427.1
29 openpose_pruned_0_3 368x368 49.9 275 44.1
30 personreid-res18_pt 176x80 1.1 275x0.9 4635.6
31 personreid-res50_pt 256x128 5.4 275x0.9 1108.6
32 plate_detection 320x320 0.49 275 5163.1
33 plate_num 96x288 1.75 275x0.9 1282.8
34 refinedet_baseline 480x360 123 275 75.9
35 RefineDet-Medical_EDD_tf 320x320 9.8 275x0.9 581.6
36 refinedet_pruned_0_8 360x480 25 275 289.1
37 refinedet_pruned_0_92 360x480 10.1 275 622.5
38 refinedet_pruned_0_96 360x480 5.1 275 903.3
39 refinedet_VOC_tf 320x320 81.9 275x0.9 97.4
40 reid 80x160 0.95 275 5219.5
41 resnet18 224x224 3.7 275 1948.1
42 resnet50 224x224 7.7 275 789.1
43 resnet50_pt 224x224 4.1 275 764.2
44 resnet50_tf2 224x224 7.7 275x0.9 711.7
45 resnet_v1_101_tf 224x224 14.4 275 460
46 resnet_v1_152_tf 224x224 21.8 275 306.7
47 resnet_v1_50_tf 224x224 7 275 885.6
48 salsanext_pt 64x2048 20.4 275x0.9 139.2
49 SemanticFPN_cityscapes_pt 256x512 10 275x0.9 585.7
50 semantic_seg_citys_tf2 512x1024 54 275x0.9 67.8
51 sp_net 128x224 0.55 275 2590.4
52 squeezenet 227x227 0.76 275 4432.2
53 squeezenet_pt 224x224 0.82 275 2817
54 ssd_adas_pruned_0_95 360x480 6.3 275 904
55 ssd_pedestrian_pruned_0_97 360x360 5.9 275x0.9 768.2
56 ssd_resnet_50_fpn_coco_tf 640x640 178.4 275x0.9 43.2
57 ssd_traffic_pruned_0_9 360x480 11.6 275 595.9
58 tiny_yolov3_vmss 416x416 5.46 275x0.9 1193.1
59 unet_chaos-CT_pt 512x512 23.3 275x0.9 81.2
60 vgg_16_tf 224x224 31 275 221.8
61 vgg_19_tf 224x224 39.3 275 184.8
62 vpgnet_pruned_0_99 480x640 2.5 275x0.9 620.7
63 yolov2_voc 448x448 34 275x0.9 225.3
64 yolov2_voc_pruned_0_66 448x448 11.6 275x0.9 567.2
65 yolov2_voc_pruned_0_71 448x448 9.9 275x0.9 665.3
66 yolov2_voc_pruned_0_77 448x448 7.8 275x0.9 807.3
67 yolov3_adas_pruned_0_9 256x512 5.5 275x0.9 869.6
68 yolov3_bdd 288x512 53.7 275x0.9 104.5
69 yolov3_voc 416x416 65.4 275x0.9 106
70 yolov3_voc_tf 416x416 65.6 275x0.9 106.3
71 yolov4_leaky_spp_m 416x416 60.1 275x0.9 112.5

U50LV Performance with 10E275 MHz DPUCAHX8H

The following table shows the throughput performance (in frames/sec or fps) for various neural network samples on U50LV Gen3x4 with DPUCAHX8H running at 10E@275 MHz.

Table 4. U50LV Performance with 10E275 MHz DPUCAHX8H
No Neural Network Input Size GOPS DPU Frequency (MHz) Performance (fps) (Multiple thread)
1 densebox_320_320 320x320 0.49 275 2455.1
2 densebox_640_360 360x640 1.1 275 1100.6
3 ENet_cityscapes_pt 512x1024 8.6 275x0.9 98.5
4 face_landmark 96x72 0.14 275 14684.5
5 face-quality 80x60 0.06 275x0.9 19370.2
6 face-quality_pt 80x60 0.06 275x0.9 19431.5
7 facerec_resnet20 112x96 3.5 275 2029.5
8 facerec-resnet20_mixed_pt 112x96 3.5 275x0.9 1851.5
9 facerec_resnet64 112x96 11 275 751.4
10 facereid-large_pt 96x96 0.5 275x0.9 10469.8
11 facereid-small_pt 80x80 0.09 275x0.9 21452.8
12 fpn 256x512 8.9 275x0.9 561.6
13 FPN_Res18_Medical_segmentation 320x320 45.3 275x0.9 140.1
14 FPN-resnet18_covid19-seg_pt 352x352 22.7 275x0.9 315.4
15 inception_resnet_v2_tf 299x299 26.4 275x0.8 209.4
16 inception_v1 224x224 3.2 275x0.9 1648.4
17 inception_v1_tf 224x224 3 275x0.9 1669.9
18 inception_v2 224x224 4 275x0.9 1319.4
19 inception_v3 299x299 11.4 275x0.8 493.6
20 inception_v3_pt 299x299 5.7 275x0.8 493.7
21 inception_v3_tf 299x299 11.5 275x0.8 493.6
22 inception_v3_tf2 299x299 11.5 275x0.9 561.4
23 inception_v4 299x299 24.5 275x0.8 226.4
24 inception_v4_2016_09_09_tf 299x299 24.6 275x0.8 226.4
25 medical_seg_cell_tf2 128x128 5.3 275x0.9 1582.5
26 MLPerf_resnet50_v1.5_tf 224x224 8.19 275x0.9 765.5
27 mlperf_ssd_resnet34_tf 1200x1200 433 275x0.9 20.7
28 multi_task 288x512 14.8 275x0.9 456.2
29 openpose_pruned_0_3 368x368 49.9 275x0.7 34.6
30 personreid-res18_pt 176x80 1.1 275x0.9 5134.6
31 personreid-res50_pt 256x128 5.4 275x0.9 1230.1
32 plate_detection 320x320 0.49 275 5252.2
33 plate_num 96x288 1.75 275x0.9 1341
34 refinedet_baseline 480x360 123 275x0.7 26
35 RefineDet-Medical_EDD_tf 320x320 9.8 275x0.9 638.1
36 refinedet_pruned_0_8 360x480 25 275x0.9 287.8
37 refinedet_pruned_0_92 360x480 10.1 275x0.9 610
38 refinedet_pruned_0_96 360x480 5.1 275x0.9 871.6
39 refinedet_VOC_tf 320x320 81.9 275x0.9 107.7
40 reid 80x160 0.95 275 5782.5
41 resnet18 224x224 3.7 275x0.9 1936.7
42 resnet50 224x224 7.7 275x0.9 790.2
43 resnet50_pt 224x224 4.1 275x0.9 765.4
44 resnet50_tf2 224x224 7.7 275x0.9 790.3
45 resnet_v1_101_tf 224x224 14.4 275x0.9 460.6
46 resnet_v1_152_tf 224x224 21.8 275x0.9 307
47 resnet_v1_50_tf 224x224 7 275x0.9 887.2
48 salsanext_pt 64x2048 20.4 275x0.9 132.7
49 SemanticFPN_cityscapes_pt 256x512 10 275x0.9 621.8
50 semantic_seg_citys_tf2 512x1024 54 275x0.9 72.9
51 sp_net 128x224 0.55 275 2745.2
52 squeezenet 227x227 0.76 275x0.9 4184.6
53 squeezenet_pt 224x224 0.82 275x0.9 2628.3
54 ssd_adas_pruned_0_95 360x480 6.3 275x0.8 807.6
55 ssd_pedestrian_pruned_0_97 360x360 5.9 275x0.9 827.9
56 ssd_resnet_50_fpn_coco_tf 640x640 178.4 275x0.8 41.9
57 ssd_traffic_pruned_0_9 360x480 11.6 275 644.9
58 tiny_yolov3_vmss 416x416 5.46 275x0.9 1309.9
59 unet_chaos-CT_pt 512x512 23.3 275x0.9 89.6
60 vgg_16_tf 224x224 31 275x0.9 223.2
61 vgg_19_tf 224x224 39.3 275x0.9 186.2
62 vpgnet_pruned_0_99 480x640 2.5 275x0.9 643
63 yolov2_voc 448x448 34 275x0.8 220.8
64 yolov2_voc_pruned_0_66 448x448 11.6 275x0.8 557.8
65 yolov2_voc_pruned_0_71 448x448 9.9 275x0.8 653.5
66 yolov2_voc_pruned_0_77 448x448 7.8 275x0.8 795.3
67 yolov3_adas_pruned_0_9 256x512 5.5 275x0.8 815.6
68 yolov3_bdd 288x512 53.7 275x0.8 101.6
69 yolov3_voc 416x416 65.4 275x0.8 102.9
70 yolov3_voc_tf 416x416 65.6 275x0.8 103.4
71 yolov4_leaky_spp_m 416x416 60.1 275x0.8 109.1

Model end-to-end Performance on U50LV 300 MHz Single Core DPUCAHX8L

The following table shows the throughput performance (in frames/sec or fps) for various neural network samples on U50LV Gen3x4 with single DPUCAHX8L running at 300 MHz.

Table 5. Model end-to-end Performance on U50LV 300 MHz Single Core DPUCAHX8L
No Neural Network Input Size GOPS Performance (fps) (Single thread) Performance (fps) (Multiple thread)
1 densebox_320_320 320x320 0.49 187.2 365.5
2 densebox_640_360 360x640 1.1 89.2 164.7
3 ENet_cityscapes_pt 512x1024 8.6 4.3 5.1
4 face_landmark 96x72 0.14 1678.7 4507.5
5 face-quality 80x60 0.06 1983.3 6735.3
6 face-quality_pt 80x60 0.06 2048.9 6828.4
7 facerec_resnet20 112x96 3.5 234.2 288.3
8 facerec-resnet20_mixed_pt 112x96 3.5 234.9 288.1
9 facerec_resnet64 112x96 11 114.0 127.4
10 facereid-small_pt 80x80 0.09 1492.9 4365.4
11 fpn 256x512 8.9 23.1 33.7
12 FPN_Res18_Medical_segmentation 320x320 45.3 12.3 13.2
13 FPN-resnet18_covid19-seg_pt 352x352 22.7 64.8 80.5
14 inception_resnet_v2_tf 299x299 26.4 27.1 30.1
15 inception_v1 224x224 3.2 202.1 284.8
16 inception_v1_tf 224x224 3 182.1 295.9
17 inception_v2 224x224 3.88 137.1 180.4
18 inception_v3 299x299 11.4 80.1 97.1
19 inception_v3_pt 299x299 5.7 72.3 99.1
20 inception_v3_tf 299x299 11.5 72.0 98.2
21 inception_v3_tf2 299x299 11.5 68.7 88.6
22 inception_v4 299x299 24.5 41.1 45.9
23 inception_v4_2016_09_09_tf 299x299 24.6 38.9 45.8
24 medical_seg_cell_tf2 128x128 5.3 70.7 75.8
25 MLPerf_resnet50_v1.5_tf 224x224 8.19 66.6 77.8
26 mlperf_ssd_resnet34_tf 1200x1200 433 3.8 6.0
27 mobilenet_1_0_224_tf2 224x224 1.1 465.0 1613.2
28 mobilenet_v1_0_5_160_tf 160x160 0.15 918.1 4718.0
29 mobilenet_v1_1_0_224_tf 224x224 1.1 473.7 1658.4
30 mobilenet_v2 224x224 0.6 431.4 984.5
31 mobilenet_v2_1_0_224_tf 224x224 0.6 377.0 955.3
32 mobilenet_v2_1_4_224_tf 224x224 1.2 306.7 702.3
33 multi_task 288x512 14.8 13.1 17.7
34 openpose_pruned_0_3 368x368 49.9 8.7 14.2
35 personreid-res50_pt 256x128 5.4 84.1 97.4
36 plate_detection 320x320 0.49 369.9 1091.1
37 refinedet_baseline 480x360 123 22.2 25.7
38 RefineDet-Medical_EDD_tf 320x320 9.8 75.1 140.1
39 refinedet_pruned_0_8 360x480 25 46.4 64.6
40 refinedet_pruned_0_92 360x480 10.1 50.9 72.5
41 refinedet_pruned_0_96 360x480 5.1 59.0 89.1
42 refinedet_VOC_tf 320x320 81.9 23.3 39.0
43 reid 80x160 0.95 378.8 542.2
44 resnet18 224x224 3.7 190.3 263.1
45 resnet50 224x224 7.7 69.5 77.2
46 resnet50_pt 224x224 4.1 66.9 78.6
47 resnet50_tf2 224x224 7.7 69.7 78.6
48 resnet_v1_101_tf 224x224 14.4 44.9 50.6
49 resnet_v1_152_tf 224x224 21.8 31.6 33.2
50 resnet_v1_50_tf 224x224 7 74.8 90.3
51 retinaface 360x640 1.11 82.5 183.9
52 salsanext_pt 64x2048 20.4 6.7 9.8
53 SemanticFPN_cityscapes_pt 256x512 10 22.9 34.0
54 semantic_seg_citys_tf2 512x1024 54 4.4 5.3
55 sp_net 128x224 0.55 1153.4 8313.6
56 squeezenet 227x227 0.76 325.0 680.3
57 squeezenet_pt 224x224 0.82 246.0 420.3
58 ssd_adas_pruned_0_95 360x480 6.3 60.7 116.7
59 ssdlite_mobilenet_v2_coco_tf 300x300 1.5 189.6 487.7
60 ssd_mobilenet_v1_coco_tf 300x300 2.5 223.3 769.8
61 ssd_mobilenet_v2 360x480 6.6 34.0 135.5
62 ssd_mobilenet_v2_coco_tf 300x300 3.8 111.5 203.1
63 ssd_pedestrian_pruned_0_97 360x360 5.9 22.8 30.2
64 ssd_traffic_pruned_0_9 360x480 11.6 49.8 106.2
65 vgg_16_tf 224x224 31 46.0 48.9
66 vgg_19_tf 224x224 39.3 41.1 44.4
67 vpgnet_pruned_0_99 480x640 2.5 17.9 20.1