KV260 视觉 AI 入门套件 - 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

KV260 入门套件使用的是 UltraScale+ 器件。它在编程逻辑中实现了单个 B4096 DPU 核，并为深度学习推断加速提供了 1.23 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 KV260 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. KV260 入门套件性能
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 3.6 4
2 c2d2_lite 512x512 6.86 3.2 3.4
3 centerpoint 2560x40x4 54 17.2 20.2
4 chen_color_resnet18_pt 224x224 3.627 219.9 239.7
5 clocs 12000x100x4 41 3 8.9
6 densebox_320_320 320x320 0.49 546.3 1033.2
7 densebox_640_360 360x640 1.1 270.4 493.8
8 drunet_pt 528x608 2.59 64.4 78.1
9 efficientdet_d2_tf 768x768 11.06 3.3 4.2
10 efficientnet-b0_tf2 224x224 0.36 84.3 87.7
11 efficientNet-edgetpu-L_tf 300x300 19.36 37.5 38.6
12 efficientNet-edgetpu-M_tf 240x240 7.34 86 90
13 efficientNet-edgetpu-S_tf 224x224 4.72 123.6 131.8
14 ENet_cityscapes_pt 512x1024 8.6 11.2 29.9
15 face_landmark 96x72 0.14 1030.4 1153.6
16 face_mask_detection_pt 512x512 0.593 122.3 170.5
17 face-quality 80x60 0.06 3392 5528.3
18 face-quality_pt 80x60 0.06 3351.2 5550.1
19 facerec_resnet20 112x96 3.5 180.1 184.7
20 facerec_resnet64 112x96 11 58.8 79.1
21 facerec-resnet20_mixed_pt 112x96 3.5 180.5 185.1
22 facereid-large_pt 96x96 0.5 1024.3 1222.4
23 facereid-small_pt 80x80 0.09 2457.4 3563.1
24 fadnet 576x960 441 1.7 2.2
25 fadnet_pruned 576x960 154 2.7 4.3
26 FairMot_pt 640x480 36 24.2 27
27 fpn 256x512 8.9 37.8 82.3
28 FPN_Res18_Medical_segmentation 320x320 45.3 13.6 17.5
29 FPN-resnet18_covid19-seg_pt 352x352 22.7 39.7 41.3
30 FPN-resnet18_Endov 240x320 13.75 39.7 67.9
31 HardNet_MSeg_pt 352x352 22.78 26.6 27.6
32 hfnet_tf 960x960 20.09 3.4 11
33 hourglass-pe_mpii 256x256 10.2 20.6 66.3
34 inception_resnet_v2_tf 299x299 26.4 25.6 26.2
35 inception_v1 224x224 3.2 190 221.8
36 inception_v1_tf 224x224 3 205 227.4
37 inception_v2 224x224 4 141.5 157.7
38 inception_v2_tf 224x224 3.88 99.3 104.5
39 inception_v3 299x299 11.4 64.1 67.5
40 inception_v3_pt 299x299 5.7 63.7 67.5
41 inception_v3_tf 299x299 11.5 64 67.4
42 inception_v3_tf2 299x299 11.5 63.3 66.7
43 inception_v4 299x299 24.5 30.7 31.6
44 inception_v4_2016_09_09_tf 299x299 24.6 30.8 31.6
45 medical_seg_cell_tf2 128x128 5.3 165.3 179.7
46 MLPerf_resnet50_v1.5_tf 224x224 8.19 85.1 88.7
47 mlperf_ssd_resnet34_tf 1200x1200 433 1.9 2.6
48 mobilenet_1_0_224_tf2 224x224 1.1 348.3 417.6
49 mobilenet_edge_0_75_tf 224x224 0.62 284.7 328.6
50 mobilenet_edge_1_0_tf 224x224 0.99 231.7 261
51 mobilenet_v1_0_25_128_tf 128x128 0.027 1461.1 2466.9
52 mobilenet_v1_0_5_160_tf 160x160 0.15 1005.3 1491.7
53 mobilenet_v1_1_0_224_tf 224x224 1.1 353.2 426.1
54 mobilenet_v2 224x224 0.6 297.1 344.8
55 mobilenet_v2_1_0_224_tf 224x224 0.6 291 337.4
56 mobilenet_v2_1_4_224_tf 224x224 1.2 204.3 229.1
57 mobilenet_v2_cityscapes_tf 1024x2048 132.74 1.9 3.3
58 mobilenet_v3_small_1_0_tf2 224x224 0.132 374.3 455
59 movenet_ntd_pt 192x192 0.5 102.9 352.3
60 MT-resnet18_mixed_pt 512x320 13.65 35.2 50.5
61 multi_task 288x512 14.8 42.6 58.5
62 multi_task_v3_pt 320x512 25.44 17.9 29.9
63 ocr_pt 960x960 875.7 1.1 1.3
64 ofa_depthwise_res50_pt 176x176 1.25 101.8 271
65 ofa_rcan_latency_pt 360x640 45.7 18 19.3
66 ofa_resnet50_0_9B_pt 160x160 0.9 199.2 214.9
67 ofa_yolo_pruned_0_30_pt 640x640 34.71 23 26.4
68 ofa_yolo_pruned_0_50_pt 640x640 24.62 29.4 35.5
69 ofa_yolo_pt 640x640 48.88 17.9 20.1
70 openpose_pruned_0_3 368x368 49.9 3.9 5.5
71 person-orientation_pruned_558m_pt 224x112 0.558 714.5 805.7
72 personreid-res18_pt 176x80 1.1 398.1 436.9
73 personreid-res50_pt 256x128 5.4 115.5 122.4
74 plate_detect 320x320 0.49 686.2 1287.4
75 plate_num 96x288 1.75 222.8 292.3
76 pmg_pt 224x224 2.28 162 173.3
77 pointpainting_nuscenes_pt 40000x64x16 112 1.3 2.8
78 pointpillars_kitti_pt 12000x100x4 10.8 22.5 32.3
79 pointpillars_nuscenes_pt 40000x64x5 108 2.4 5.4
80 rcan_pruned_tf 360x640 86.95 9.2 9.5
81 refinedet_baseline 480x360 123 6 9.3
82 refinedet_pruned_0_8 360x480 25 35.2 37.7
83 refinedet_pruned_0_92 360x480 10.1 70.3 80
84 refinedet_pruned_0_96 360x480 5.1 99.2 118.7
85 refinedet_VOC_tf 320x320 81.9 11 13.1
86 RefineDet-Medical_EDD_tf 320x320 9.8 72.6 85.7
87 reid 80x160 0.95 364.1 450.1
88 resnet_v1_101_tf 224x224 14.4 49.1 51.1
89 resnet_v1_152_tf 224x224 21.8 34 34.5
90 resnet_v1_50_tf 224x224 7 94.7 99.3
91 resnet_v2_101_tf 299x299 26.78 25.4 25.9
92 resnet_v2_152_tf 299x299 40.47 17.2 17.5
93 resnet_v2_50_tf 299x299 13.1 48.5 50.5
94 resnet18 224x224 3.7 211.7 234.9
95 resnet50 224x224 7.7 95.3 99.7
96 resnet50_pt 224x224 4.1 84.4 88
97 resnet50_tf2 224x224 7.7 93.4 98.2
98 retinaface 360x640 1.11 152.3 315.8
99 SA_gate_base_pt 360x360 178 3.5 4.6
100 salsanext_pt 64x2048 20.4 6.2 19.6
101 salsanext_v2_pt 64x2048 32 4.5 11.4
102 semantic_seg_citys_tf2 512x1024 54 8 15.4
103 SemanticFPN_cityscapes_pt 256x512 10 38.7 86.6
104 SemanticFPN_Mobilenetv2_pt 512x1024 5.4 11.2 33
105 SESR_S_pt 360x640 7.48 94.8 105.6
106 solo_pt 640x640 107 1.4 4.4
107 sp_net 128x224 0.55 637.1 783.8
108 squeezenet 227x227 0.76 596.2 725.3
109 squeezenet_pt 224x224 0.82 533.7 755.9
110 ssd_adas_pruned_0_95 360x480 6.3 98.4 123.4
111 ssd_inception_v2_coco_tf 300x300 9.6 42.2 47.3
112 ssd_mobilenet_v1_coco_tf 300x300 2.5 117.1 169.7
113 ssd_mobilenet_v2 360x480 6.6 26.9 66.1
114 ssd_mobilenet_v2_coco_tf 300x300 3.8 87.9 113.1
115 ssd_pedestrian_pruned_0_97 360x360 5.9 86.8 110.5
116 ssd_resnet_50_fpn_coco_tf 640x640 178.4 2.9 5.2
117 ssd_traffic_pruned_0_9 360x480 11.6 60.5 76.3
118 ssdlite_mobilenet_v2_coco_tf 300x300 1.5 106.3 161.6
119 ssr_pt 256x256 39.72 6.5 6.5
120 superpoint_tf 480x640 52.4 11.5 20.6
121 textmountain_pt 960x960 575.2 1.7 1.9
122 tiny_yolov3_vmss 416x416 5.46 130.9 167.2
123 tsd_yolox_pt 640x640 73 14 14.6
124 ultrafast_pt 288x800 8.4 37.4 41.3
125 unet_chaos-CT_pt 512x512 23.3 24.3 30.8
126 vehicle_make_resnet18_pt 224x224 3.627 217.8 239
127 vehicle_type_resnet18_pt 224x224 3.627 220.1 240.1
128 vgg_16_tf 224x224 31 21.5 21.7
129 vgg_19_tf 224x224 39.3 18.5 18.7
130 vpgnet_pruned_0_99 480x640 2.5 111.3 162.6
131 yolov2_voc 448x448 34 28.4 29.6
132 yolov2_voc_pruned_0_66 448x448 11.6 69.8 77.8
133 yolov2_voc_pruned_0_71 448x448 9.9 81.2 91.8
134 yolov2_voc_pruned_0_77 448x448 7.8 95.7 110.7
135 yolov3_adas_pruned_0_9 256x512 5.5 105.9 136.4
136 yolov3_bdd 288x512 53.7 13.8 14.3
137 yolov3_coco_416_tf2 416x416 65.9 14.1 14.8
138 yolov3_voc 416x416 65.4 14.2 14.6
139 yolov3_voc_tf 416x416 65.6 14.4 14.9
140 yolov4_leaky_416_tf 416x416 60.3 14.4 15.4
141 yolov4_leaky_512_tf 512x512 91.2 11 11.7
142 yolov4_leaky_spp_m 416x416 60.1 14.6 15.6
143 yolov4_leaky_spp_m_pruned_0_36 416x416 38.2 20.3 22