ZCU102 Performance - 1.3 English

Vitis AI Library User Guide (UG1354)

Document ID
UG1354
Release Date
2021-02-03
Version
1.3 English

The ZCU102 evaluation board uses the mid-range ZU9 UltraScale+™ device. There are two different hardware versions of ZCU102 board, one with the serial number 0432055-04 as the header, and the other with the serial number 0432055-05 as the header. The performance of the Vitis AI Library varies between the two hardware versions (because of different DDR performance). Because the 0432055-04 version of ZCU102 has been discontinued, the following table only shows the performance of ZCU102 (0432055-05). In ZCU102 board, triple B4096F DPU cores are implemented in program logic.

Refer to the following table for throughput performance (in frames/sec or fps) for various neural network samples on ZCU102 (0432055-05) with DPU running at 281 MHz.

Note: The DPU on the ZCU102 has hardware softmax acceleration module. Due to the limitation of hardware softmax module, the software softmax is faster when the number of categories reaches 1000. Set XLNX_ENABLE_C_SOFTMAX=1 to enable the software softmax: softmax_c. The default value of XLNX_ENABLE_C_SOFTMAX is 0, which means the softmax method is selected according to the following priorities.
  1. Neon Acceleration
  2. Hardware Softmax
  3. Software Softmax_c
For ZCU102, use the following command to test the performance of classification.
env XLNX_ENABLE_C_SOFTMAX=1 ./test_performance_classification resnet50 test_performance_classification.list -t 8 -s 60
Table 1. ZCU102 (0432055-05) Performance
No Neural Network Input Size GOPS Performance (fps) (Single thread) Performance (fps) (Multiple thread)
1 densebox_320_320 320x320 0.49 428.9 1638.6
2 densebox_640_360 360x640 1.1 220.1 889.6
3 ENet_cityscapes_pt 512x1024 8.6 9.1 38.5
4 face_landmark 96x72 0.14 846.1 1560.7
5 face-quality 80x60 0.06 2133.2 6681.4
6 face-quality_pt 80x60 0.06 2117.3 7410.8
7 facerec_resnet20 112x96 3.5 163.5 329
8 facerec-resnet20_mixed_pt 112x96 3.5 165.1 330.7
9 facerec_resnet64 112x96 11 71.2 177.5
10 facereid-large_pt 96x96 0.5 844.3 2177.7
11 facereid-small_pt 80x80 0.09 1869.9 5994.9
12 fpn 256x512 8.9 33.3 148.4
13 FPN_Res18_Medical_segmentation 320x320 45.3 12.3 44.9
14 FPN-resnet18_covid19-seg_pt 352x352 22.7 36 107
15 FPN-resnet18_Endov 240x320 13.75 33.6 160.2
16 hourglass-pe_mpii 256x256 10.2 18.2 76.1
17 inception_resnet_v2_tf 299x299 26.4 22.2 50
18 inception_v1 224x224 3.2 169.1 429.3
19 inception_v1_tf 224x224 3 172.4 429.6
20 inception_v2 224x224 4 125.2 290
21 inception_v2_tf 224x224 3.88 88.1 222.7
22 inception_v3 299x299 11.4 58.5 135.7
23 inception_v3_pt 299x299 5.7 58.6 136.1
24 inception_v3_tf 299x299 11.5 58 134.1
25 inception_v3_tf2 299x299 11.5 57.1 133.7
26 inception_v4 299x299 24.5 28.6 68.8
27 inception_v4_2016_09_09_tf 299x299 24.6 28.5 68.7
28 medical_seg_cell_tf2 128x128 5.3 154.1 392.6
29 MLPerf_resnet50_v1.5_tf 224x224 8.19 71.2 167.9
30 mlperf_ssd_resnet34_tf 1200x1200 433 2 7
31 mobilenet_1_0_224_tf2 224x224 1.1 292.6 918.9
32 mobilenet_edge_0_75_tf 224x224 0.62 237.6 699.4
33 mobilenet_edge_1_0_tf 224x224 0.99 198.3 544.3
34 mobilenet_v1_0_25_128_tf 128x128 0.027 1127 3929.4
35 mobilenet_v1_0_5_160_tf 160x160 0.15 714.6 2692.3
36 mobilenet_v1_1_0_224_tf 224x224 1.1 297.3 931.4
37 mobilenet_v2 224x224 0.6 250 729.2
38 mobilenet_v2_1_0_224_tf 224x224 0.6 245 689.9
39 mobilenet_v2_1_4_224_tf 224x224 1.2 179.7 467.8
40 mobilenet_v2_cityscapes_tf 1024x2048 132.74 1.6 4.6
41 MT-resnet18_mixed_pt 512x320 13.65 29.3 100.2
42 multi_task 288x512 14.8 35.2 125.2
43 openpose_pruned_0_3 368x368 49.9 3.5 15
44 personreid-res18_pt 176x80 1.1 350 682.2
45 personreid-res50_pt 256x128 5.4 96.9 227.3
46 plate_detection 320x320 0.49 505.9 2025.7
47 plate_num 96x288 1.75 149.8 441.1
48 pointpillars_kitti_12000_0_pt pointpillars_kitti_12000_1_pt 12000x100 10.8 19.9 49.9
49 refinedet_baseline 480x360 123 8.5 24.7
50 RefineDet-Medical_EDD_tf 320x320 9.8 66.1 225
51 refinedet_pruned_0_8 360x480 25 32.6 100.3
52 refinedet_pruned_0_92 360x480 10.1 63.1 201.8
53 refinedet_pruned_0_96 360x480 5.1 87.4 288.2
54 refinedet_VOC_tf 320x320 81.9 11.3 34.4
55 reid 80x160 0.95 351 689.8
56 resnet18 224x224 3.7 183.2 476.4
57 resnet50 224x224 7.7 72.7 173.1
58 resnet50_pt 224x224 4.1 68.8 165.8
59 resnet50_tf2 224x224 7.7 70.7 170.5
60 resnet_v1_101_tf 224x224 14.4 42.6 106.2
61 resnet_v1_152_tf 224x224 21.8 29.1 73.9
62 resnet_v1_50_tf 224x224 7 79.6 184.8
63 resnet_v2_101_tf 299x299 26.78 20.5 53.8
64 resnet_v2_152_tf 299x299 40.47 14.5 37.2
65 resnet_v2_50_tf 299x299 13.1 35.1 93.8
66 retinaface 360x640 1.11 126.2 578.2
67 salsanext_pt 64x2048 20.4 5.3 19.7
68 SemanticFPN_cityscapes_pt 256x512 10 33.2 161.3
69 semantic_seg_citys_tf2 512x1024 54 6.8 23.9
70 sp_net 128x224 0.55 485.9 1408.8
71 squeezenet 227x227 0.76 270.6 1119.8
72 squeezenet_pt 224x224 0.82 221.7 847.2
73 ssd_adas_pruned_0_95 360x480 6.3 87 298.1
74 ssd_inception_v2_coco_tf 300x300 9.6 38.5 101
75 ssdlite_mobilenet_v2_coco_tf 300x300 1.5 101.7 303.9
76 ssd_mobilenet_v1_coco_tf 300x300 2.5 107.9 324.4
77 ssd_mobilenet_v2 360x480 6.6 39.1 115.9
78 ssd_mobilenet_v2_coco_tf 300x300 3.8 79.7 210.9
79 ssd_pedestrian_pruned_0_97 360x360 5.9 76.3 279.2
80 ssd_resnet_50_fpn_coco_tf 640x640 178.4 2.9 5.2
81 ssd_traffic_pruned_0_9 360x480 11.6 54.9 200.2
82 tiny_yolov3_vmss 416x416 5.46 117.7 390.1
83 unet_chaos-CT_pt 512x512 23.3 21.2 68.1
84 vgg_16_tf 224x224 31 20.1 41
85 vgg_19_tf 224x224 39.3 17.3 36.5
86 vpgnet_pruned_0_99 480x640 2.5 96.9 360.3
87 yolov2_voc 448x448 34 26.4 69.6
88 yolov2_voc_pruned_0_66 448x448 11.6 62.8 189.1
89 yolov2_voc_pruned_0_71 448x448 9.9 72.2 220.6
90 yolov2_voc_pruned_0_77 448x448 7.8 84.3 265
91 yolov3_adas_pruned_0_9 256x512 5.5 88.3 260.7
92 yolov3_bdd 288x512 53.7 12.4 32.8
93 yolov3_voc 416x416 65.4 12.7 33.4
94 yolov3_voc_tf 416x416 65.6 13.2 34.3
95 yolov4_leaky_spp_m 416x416 60.1 13.1 33.5