ZCU102 Performance - 1.2 English

Vitis AI Library User Guide (UG1354)
Document ID
UG1354
Release Date
2020-07-21
Version
1.2 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). Since 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 will be 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 will be selected according to the following priorities.
  1. Neon Acceleration
  2. Hardware Softmax
  3. Software Softmax_c

For ZCU102, you can 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 inception_resnet_v2_tf 299x299 26.4 23.1 48.7
2 inception_v1_tf 224x224 3.0 184.1 423.9
3 inception_v3_tf 299x299 11.5 57.3 126.7
4 inception_v4_2016_09_09_tf 299x299 24.6 28.5 66.2
5 mobilenet_v1_0_25_128_tf 128x128 0.027 1170.7 4043.5
6 mobilenet_v1_0_5_160_tf 160x160 0.15 707.6 2007.1
7 mobilenet_v1_1_0_224_tf 224x224 1.1 284.3 754.9
8 mobilenet_v2_1_0_224_tf 224x224 0.60 230.8 568.4
9 mobilenet_v2_1_4_224_tf 224x224 1.2 167.3 393.1
10 resnet_v1_101_tf 224x224 14.4 43.1 91.3
11 resnet_v1_152_tf 224x224 21.8 29.6 63.7
12 resnet_v1_50_tf 224x224 7.0 79.1 161.9
13 vgg_16_tf 224x224 31.0 20.1 40.9
14 vgg_19_tf 224x224 39.3 17.3 36.5
15 ssd_mobilenet_v1_coco_tf 300x300 2.5 90.1 332.9
16 ssd_mobilenet_v2_coco_tf 300x300 3.8 63.9 193.2
17 ssd_resnet_50_fpn_coco_tf 640x640 178.4 1.3 5.1
18 yolov3_voc_tf 416x416 65.6 13.5 35
19 mlperf_ssd_resnet34_tf 1200x1200 433 2 7.2
20 resnet50 224x224 7.7 73.5 152.7
21 resnet18 224x224 3.7 186.9 441.6
22 inception_v1 224x224 3.2 178 411.7
23 inception_v2 224x224 4.0 144.4 317.3
24 inception_v3 299x299 11.4 57.5 128.1
25 inception_v4 299x299 24.5 28.5 66.2
26 mobilenet_v2 224x224 0.6 226.8 548
27 squeezenet 227x227 0.76 265.8 1012.3
28 ssd_pedestrain_pruned_0_97 360x360 5.9 76.3 282.6
29 ssd_traffic_pruned_0_9 360x480 11.6 54.5 201.5
30 ssd_adas_pruned_0_95 360x480 6.3 82.9 279.7
31 ssd_mobilenet_v2 360x480 6.6 38.4 114.4
32 refinedet_pruned_0_8 360x480 25 31.7 101.3
33 refinedet_pruned_0_92 360x480 10.1 59.9 196.8
34 refinedet_pruned_0_96 360x480 5.1 82.9 276.2
35 vpgnet_pruned_0_99 480x640 2.5 104.5 381.4
36 fpn 256x512 8.9 59.7 175.5
37 sp_net 128x224 0.55 381.6 1317.4
38 openpose_pruned_0_3 368x368 49.9 3.5 15.1
39 densebox_320_320 320x320 0.49 390 1172.3
40 densebox_640_360 360x640 1.1 200.4 588.7
41 face_landmark 96x72 0.14 849.4 1382.7
42 reid 80x160 0.95 364.2 665.6
43 multi_task 288x512 14.8 35.5 127.7
44 yolov3_adas_pruned_0_9 256x512 5.5 84.1 229.7
45 yolov3_voc 416x416 65.4 13.5 35.3
46 yolov3_bdd 288x512 53.7 13 34.3
47 yolov2_voc 448x448 34 26.8 71
48 yolov2_voc_pruned_0_66 448x448 11.6 63.2 185.9
49 yolov2_voc_pruned_0_71 448x448 9.9 72.8 214.8
50 yolov2_voc_pruned_0_77 448x448 7.8 85.2 258.7
51 facerec_resnet20 112x96 3.5 167.1 320.6
52 facerec_resnet64 112x96 11.0 73 173
53 plate_detection 320x320 0.49 500 1792.2
54 plate_recognition 96x288 1.75 113.4 383.2
55 FPN_Res18_Medical_segmentation 320x320 45.3 12.2 40.3
56 refinedet_baseline 480x360 123 8.3 24.4