ZCU104 评估板 - 2.5 简体中文

Vitis AI Library 用户指南 (UG1354)
Document ID
UG1354
Release Date
2022-06-15
Version
2.5 简体中文

ZCU104 评估板使用中端 ZU7ev UltraScale+ 器件。它在编程逻辑区域实现了双 B4096 DPU 核，并为深度学习推断加速提供了 2.46 TOPS INT8 峰值性能。

请参阅下表，查看具有运行频率为 300 MHz 的 DPU 的 ZCU104 上的各种神经网络采样的吞吐量性能（以每秒帧数或 fps 为单位）。

表 1. ZCU104 性能
编号 神经网络 输入大小 GOPS 性能 (fps)（单线程） 性能 (fps)（多线程）
1 bcc_pt 800x1000 268.9 3.5 8
2 c2d2_lite 512x512 6.86 3.1 3.5
3 centerpoint 2560x40x4 54 16.8 20.7
4 chen_color_resnet18_pt 224x224 3.627 217.7 437.1
5 clocs 12000x100x4 41 2.9 10.3
6 densebox_320_320 320x320 0.49 513.4 1697.4
7 densebox_640_360 360x640 1.1 255.4 811.3
8 drunet_pt 528x608 2.59 63.8 152.6
9 efficientdet_d2_tf 768x768 11.06 3.3 6.3
10 efficientnet-b0_tf2 224x224 0.36 83.7 146.1
11 efficientNet-edgetpu-L_tf 300x300 19.36 37.3 73.4
12 efficientNet-edgetpu-M_tf 240x240 7.34 85.3 168.5
13 efficientNet-edgetpu-S_tf 224x224 4.72 122.9 249
14 ENet_cityscapes_pt 512x1024 8.6 10.4 39.4
15 face_landmark 96x72 0.14 1014.4 1669.5
16 face_mask_detection_pt 512x512 0.593 120.7 326.8
17 face-quality 80x60 0.06 3100.6 8601.7
18 face-quality_pt 80x60 0.06 3035 8512.2
19 facerec_resnet20 112x96 3.5 179.7 314
20 facerec_resnet64 112x96 11 78.2 149.6
21 facerec-resnet20_mixed_pt 112x96 3.5 180.3 315.3
22 facereid-large_pt 96x96 0.5 992.6 2122.4
23 facereid-small_pt 80x80 0.09 2306.3 5904.6
24 fadnet 576x960 441 1.7 3.5
25 fadnet_pruned 576x960 154 2.6 5.3
26 FairMot_pt 640x480 36 23.8 52.4
27 fpn 256x512 8.9 35.6 142.1
28 FPN_Res18_Medical_segmentation 320x320 45.3 13.3 34.7
29 FPN-resnet18_covid19-seg_pt 352x352 22.7 39.4 81
30 FPN-resnet18_Endov 240x320 13.75 38.7 130.3
31 HardNet_MSeg_pt 352x352 22.78 26.4 49.4
32 hfnet_tf 960x960 20.09 3.2 14.3
33 hourglass-pe_mpii 256x256 10.2 19.2 71.5
34 inception_resnet_v2_tf 299x299 26.4 25.5 47
35 inception_v1 224x224 3.2 198.7 406.6
36 inception_v1_tf 224x224 3 202.4 412.9
37 inception_v2 224x224 4 145.1 278.4
38 inception_v2_tf 224x224 3.88 99 195.2
39 inception_v3 299x299 11.4 63.6 122
40 inception_v3_pt 299x299 5.7 63.5 122.1
41 inception_v3_tf 299x299 11.5 63.4 121.5
42 inception_v3_tf2 299x299 11.5 62.9 121.6
43 inception_v4 299x299 24.5 30.7 58.9
44 inception_v4_2016_09_09_tf 299x299 24.6 30.7 59
45 medical_seg_cell_tf2 128x128 5.3 163.5 338.6
46 MLPerf_resnet50_v1.5_tf 224x224 8.19 84.2 157.7
47 mlperf_ssd_resnet34_tf 1200x1200 433 1.9 5.2
48 mobilenet_1_0_224_tf2 224x224 1.1 339 772.4
49 mobilenet_edge_0_75_tf 224x224 0.62 278.4 602.4
50 mobilenet_edge_1_0_tf 224x224 0.99 227.5 475.1
51 mobilenet_v1_0_25_128_tf 128x128 0.027 1370.4 4280.9
52 mobilenet_v1_0_5_160_tf 160x160 0.15 948.6 2639.2
53 mobilenet_v1_1_0_224_tf 224x224 1.1 344.1 782
54 mobilenet_v2 224x224 0.6 291 631
55 mobilenet_v2_1_0_224_tf 224x224 0.6 284.3 609.1
56 mobilenet_v2_1_4_224_tf 224x224 1.2 202.7 412.2
57 mobilenet_v2_cityscapes_tf 1024x2048 132.74 1.8 5.5
58 mobilenet_v3_small_1_0_tf2 224x224 0.132 363.6 819
59 movenet_ntd_pt 192x192 0.5 95.4 366.4
60 MT-resnet18_mixed_pt 512x320 13.65 34.3 93.1
61 multi_task 288x512 14.8 41.3 111.2
62 multi_task_v3_pt 320x512 25.44 17.8 55.4
63 ocr_pt 960x960 875.7 1 2.6
64 ofa_depthwise_res50_pt 176x176 1.25 109.5 342.7
65 ofa_rcan_latency_pt 360x640 45.7 17.9 28.8
66 ofa_resnet50_0_9B_pt 160x160 0.9 191.5 354
67 ofa_yolo_pruned_0_30_pt 640x640 34.71 22.8 48.7
68 ofa_yolo_pruned_0_50_pt 640x640 24.62 29.2 64.5
69 ofa_yolo_pt 640x640 48.88 17.9 37.4
70 openpose_pruned_0_3 368x368 49.9 4 11
71 person-orientation_pruned_558m_pt 224x112 0.558 700.7 1368.5
72 personreid-res18_pt 176x80 1.1 395.3 700.2
73 personreid-res50_pt 256x128 5.4 114.6 216
74 plate_detect 320x320 0.49 642.8 2215
75 plate_num 96x288 1.75 211.6 485.3
76 pmg_pt 224x224 2.28 161.5 318.7
77 pointpainting_nuscenes_pt 40000x64x16 112 1.3 4.5
78 pointpillars_kitti_pt 12000x100x4 10.8 20.1 49.6
79 pointpillars_nuscenes_pt 40000x64x5 108 2.3 8.9
80 rcan_pruned_tf 360x640 86.95 9.2 17.1
81 refinedet_baseline 480x360 123 9.1 18.5
82 refinedet_pruned_0_8 360x480 25 35 73.9
83 refinedet_pruned_0_92 360x480 10.1 69.3 154.9
84 refinedet_pruned_0_96 360x480 5.1 97.2 227
85 refinedet_VOC_tf 320x320 81.9 10.9 25.9
86 RefineDet-Medical_EDD_tf 320x320 9.8 71.3 169.6
87 reid 80x160 0.95 395.3 715.2
88 resnet_v1_101_tf 224x224 14.4 49.5 94.7
89 resnet_v1_152_tf 224x224 21.8 33.7 64.8
90 resnet_v1_50_tf 224x224 7 93.9 175.2
91 resnet_v2_101_tf 299x299 26.78 25.2 47.9
92 resnet_v2_152_tf 299x299 40.47 17.2 32.6
93 resnet_v2_50_tf 299x299 13.1 48.1 90.4
94 resnet18 224x224 3.7 208.8 428.1
95 resnet50 224x224 7.7 92.5 176.6
96 resnet50_pt 224x224 4.1 83.6 157.1
97 resnet50_tf2 224x224 7.7 93.1 175
98 retinaface 360x640 1.11 143.7 512.1
99 SA_gate_base_pt 360x360 178 3.5 8.5
100 salsanext_pt 64x2048 20.4 5.7 21.6
101 salsanext_v2_pt 64x2048 32 4.3 11.7
102 semantic_seg_citys_tf2 512x1024 54 7.7 25.6
103 SemanticFPN_cityscapes_pt 256x512 10 36.4 149.7
104 SemanticFPN_Mobilenetv2_pt 512x1024 5.4 10.7 52
105 SESR_S_pt 360x640 7.48 94.6 146
106 solo_pt 640x640 107 1.5 4.8
107 sp_net 128x224 0.55 612.3 1357.5
108 squeezenet 227x227 0.76 575.3 1252.9
109 squeezenet_pt 224x224 0.82 599.1 1317.2
110 ssd_adas_pruned_0_95 360x480 6.3 96.1 238.2
111 ssd_inception_v2_coco_tf 300x300 9.6 41.6 87.9
112 ssd_mobilenet_v1_coco_tf 300x300 2.5 115.2 307.1
113 ssd_mobilenet_v2 360x480 6.6 25.5 105
114 ssd_mobilenet_v2_coco_tf 300x300 3.8 85.6 196.5
115 ssd_pedestrian_pruned_0_97 360x360 5.9 84.6 214.4
116 ssd_resnet_50_fpn_coco_tf 640x640 178.4 2.9 5.2
117 ssd_traffic_pruned_0_9 360x480 11.6 59.2 150.1
118 ssdlite_mobilenet_v2_coco_tf 300x300 1.5 110.2 279.5
119 ssr_pt 256x256 39.72 6.5 12.4
120 superpoint_tf 480x640 52.4 11.1 40.3
121 textmountain_pt 960x960 575.2 1.8 3.7
122 tiny_yolov3_vmss 416x416 5.46 132.2 328.5
123 tsd_yolox_pt 640x640 73 14 27.8
124 ultrafast_pt 288x800 8.4 37.2 78.4
125 unet_chaos-CT_pt 512x512 23.3 23.6 59.3
126 vehicle_make_resnet18_pt 224x224 3.627 216.2 436
127 vehicle_type_resnet18_pt 224x224 3.627 218 437.2
128 vgg_16_tf 224x224 31 21.5 37
129 vgg_19_tf 224x224 39.3 18.5 32.7
130 vpgnet_pruned_0_99 480x640 2.5 108.9 301
131 yolov2_voc 448x448 34 28.3 57.1
132 yolov2_voc_pruned_0_66 448x448 11.6 69.3 153.5
133 yolov2_voc_pruned_0_71 448x448 9.9 80.1 180.7
134 yolov2_voc_pruned_0_77 448x448 7.8 93.9 217.3
135 yolov3_adas_pruned_0_9 256x512 5.5 103.2 245.6
136 yolov3_bdd 288x512 53.7 13.8 27.7
137 yolov3_coco_416_tf2 416x416 65.9 14.1 28.7
138 yolov3_voc 416x416 65.4 14.2 28.5
139 yolov3_voc_tf 416x416 65.6 14.4 28.8
140 yolov4_leaky_416_tf 416x416 60.3 14.4 28.9
141 yolov4_leaky_512_tf 512x512 91.2 10.9 21.8
142 yolov4_leaky_spp_m 416x416 60.1 14.5 29.2
143 yolov4_leaky_spp_m_pruned_0_36 416x416 38.2 20.1 40.5