VCK190 評価ボード - 2.5 日本語

Vitis AI ライブラリ ユーザー ガイド (UG1354)
Document ID
UG1354
Release Date
2022-06-15
Version
2.5 日本語

VCK190 は、Versal AI コア シリーズの最初の評価キットであり、現在のサーバー クラス CPU の 100 倍以上の計算性能を達成できる AI および DSP エンジンを活用したソリューションを開発できます。このリリースでは、AI エンジンを使用して 1 個の C32B6 DPU コアが実装されており、61.4 TOPS INT8 のピーク性能を達成できるため、深層学習の推論高速化に最適です。

次の表では、AI エンジン (動作周波数 1250MHz) と PL (動作周波数 333MHz) を使用した VCK190 のさまざまなニューラル ネットワーク サンプルのスループット性能 (fps) を示しています。

表 1. VCK190 の性能 (バッチ数 = 6)
番号 ニューラル ネットワーク 入力サイズ GOPS 性能 (fps) (単一スレッド) 性能 (fps) (複数スレッド)
1 bcc_pt 800x1000 268.9 36.9 72.4
2 c2d2_lite 512x512 6.86 20.4 26.3
3 centerpoint 2560x40x4 54 128 237.8
4 chen_color_resnet18_pt 224x224 3.627 2046.3 5223.5
5 clocs 12000x100x4 41 8.1 14.4
6 densebox_320_320 320x320 0.49 1286.9 2570.9
7 densebox_640_360 360x640 1.1 629.6 1204.2
8 drunet_pt 528x608 2.59 256 462.1
9 efficientdet_d2_tf 768x768 11.06 18.1 40.1
10 efficientnet-b0_tf2 224x224 0.36 1088.6 1812.4
11 efficientNet-edgetpu-L_tf 300x300 19.36 397.3 512.1
12 efficientNet-edgetpu-M_tf 240x240 7.34 860.4 1397
13 efficientNet-edgetpu-S_tf 224x224 4.72 1210.3 2196.1
14 ENet_cityscapes_pt 512x1024 8.6 23.6 54.4
15 face_landmark 96x72 0.14 9846.8 24605.3
16 face_mask_detection_pt 512x512 0.593 453.8 923.1
17 face-quality 80x60 0.06 13848.3 29954.4
18 face-quality_pt 80x60 0.06 13791.3 29820.2
19 facerec_resnet20 112x96 3.5 3865.1 5695.3
20 facerec_resnet64 112x96 11 2185.2 2673
21 facerec-resnet20_mixed_pt 112x96 3.5 3855 5693.6
22 facereid-large_pt 96x96 0.5 7564 19131.6
23 facereid-small_pt 80x80 0.09 11372.4 25915.7
24 fadnet 576x960 441 8 13.6
25 fadnet_pruned 576x960 154 8.4 15.9
26 FairMot_pt 640x480 36 194.7 365
27 fpn 256x512 8.9 103.9 221.3
28 FPN_Res18_Medical_segmentation 320x320 45.3 74.3 195.2
29 FPN-resnet18_covid19-seg_pt 352x352 22.7 480.8 784.7
30 FPN-resnet18_Endov 240x320 13.75 81.8 170.1
31 HardNet_MSeg_pt 352x352 22.78 219.4 274
32 hfnet_tf 960x960 20.09 9.3 22.4
33 hourglass-pe_mpii 256x256 10.2 78 133.8
34 inception_resnet_v2_tf 299x299 26.4 387.6 500.9
35 inception_v1 224x224 3.2 1308 2432.5
36 inception_v1_tf 224x224 3 1305.5 2462.2
37 inception_v2 224x224 4 1100.2 1845.8
38 inception_v2_tf 224x224 3.88 829.2 1196.5
39 inception_v3 299x299 11.4 590.9 899
40 inception_v3_pt 299x299 5.7 594.2 897.7
41 inception_v3_tf 299x299 11.5 597 905.1
42 inception_v3_tf2 299x299 11.5 657.4 1062.1
43 inception_v4 299x299 24.5 340.5 420.1
44 inception_v4_2016_09_09_tf 299x299 24.6 342.1 423.9
45 medical_seg_cell_tf2 128x128 5.3 1536 3036.3
46 MLPerf_resnet50_v1.5_tf 224x224 8.19 1367.7 2744.2
47 mlperf_ssd_resnet34_tf 1200x1200 433 11.9 20.9
48 mobilenet_1_0_224_tf2 224x224 1.1 2031.1 5024.9
49 mobilenet_edge_0_75_tf 224x224 0.62 1902.8 4940.2
50 mobilenet_edge_1_0_tf 224x224 0.99 1835.3 4746.8
51 mobilenet_v1_0_25_128_tf 128x128 0.027 5085.4 10391.8
52 mobilenet_v1_0_5_160_tf 160x160 0.15 3512.5 7804.8
53 mobilenet_v1_1_0_224_tf 224x224 1.1 2015.4 4859.5
54 mobilenet_v2 224x224 0.6 1928.2 4954.6
55 mobilenet_v2_1_0_224_tf 224x224 0.6 1899.3 4996.3
56 mobilenet_v2_1_4_224_tf 224x224 1.2 1538.8 4186.2
57 mobilenet_v2_cityscapes_tf 1024x2048 132.74 4.9 11.9
58 mobilenet_v3_small_1_0_tf2 224x224 0.132 2041.2 4910.7
59 movenet_ntd_pt 192x192 0.5 241.8 443
60 MT-resnet18_mixed_pt 512x320 13.65 131.7 257.5
61 multi_task 288x512 14.8 161.9 288
62 multi_task_v3_pt 320x512 25.44 77.8 174.6
63 ocr_pt 960x960 875.7 8.6 18.5
64 ofa_depthwise_res50_pt 176x176 1.25 302.9 450
65 ofa_rcan_latency_pt 360x640 45.7 59 81.5
66 ofa_resnet50_0_9B_pt 160x160 0.9 2090.9 4009.4
67 ofa_yolo_pruned_0_30_pt 640x640 34.71 145.8 255
68 ofa_yolo_pruned_0_50_pt 640x640 24.62 157.4 297
69 ofa_yolo_pt 640x640 48.88 127.7 220.7
70 openpose_pruned_0_3 368x368 49.9 23.5 35.4
71 person-orientation_pruned_558m_pt 224x112 0.558 5545.9 12380.2
72 personreid-res18_pt 176x80 1.1 4237.7 8631.8
73 personreid-res50_pt 256x128 5.4 1824.6 3789.4
74 plate_detect 320x320 0.49 1550.6 2969.4
75 plate_num 96x288 1.75 1158.7 2175.9
76 pmg_pt 224x224 2.28 1747.6 3627.9
77 pointpainting_nuscenes_pt 40000x64x16 112 3.8 6.6
78 pointpillars_kitti_pt 12000x100x4 10.8 25.3 34.8
79 pointpillars_nuscenes_pt 40000x64x5 108 7.7 16
80 psmnet 576x960 696 0.4 0.7
81 rcan_pruned_tf 360x640 86.95 46 56.9
82 refinedet_baseline 480x360 123 178.8 239.2
83 refinedet_pruned_0_8 360x480 25 317.6 617.6
84 refinedet_pruned_0_92 360x480 10.1 431.8 846.4
85 refinedet_pruned_0_96 360x480 5.1 502.2 915.1
86 refinedet_VOC_tf 320x320 81.9 105.1 226.4
87 RefineDet-Medical_EDD_tf 320x320 9.8 550.7 1283.6
88 reid 80x160 0.95 3998.5 8620
89 resnet_v1_101_tf 224x224 14.4 1064 1756.1
90 resnet_v1_152_tf 224x224 21.8 839 1216.8
91 resnet_v1_50_tf 224x224 7 1452.1 3067.5
92 resnet_v2_101_tf 299x299 26.78 412.8 537.9
93 resnet_v2_152_tf 299x299 40.47 330.9 406.6
94 resnet_v2_50_tf 299x299 13.1 548.6 792.9
95 resnet18 224x224 3.7 1808.3 4498.9
96 resnet50 224x224 7.7 1466.8 3079.1
97 resnet50_pt 224x224 4.1 1372.1 2772.1
98 resnet50_tf2 224x224 7.7 1463.5 3111.4
99 retinaface 360x640 1.11 408.1 835.6
100 SA_gate_base_pt 360x360 178 7.6 9.7
101 salsanext_pt 64x2048 20.4 12.3 23.9
102 salsanext_v2_pt 64x2048 32 10.6 22.3
103 semantic_seg_citys_tf2 512x1024 54 20.5 47.8
104 SemanticFPN_cityscapes_pt 256x512 10 110.5 218.2
105 SemanticFPN_Mobilenetv2_pt 512x1024 5.4 27.2 55.9
106 SESR_S_pt 360x640 7.48 339.8 634.9
107 solo_pt 640x640 107 4.3 7.3
108 sp_net 128x224 0.55 2704 5816.5
109 squeezenet 227x227 0.76 3144.1 5625.9
110 squeezenet_pt 224x224 0.82 3165.4 5758.6
111 ssd_adas_pruned_0_95 360x480 6.3 454.6 825.6
112 ssd_inception_v2_coco_tf 300x300 9.6 266.3 415.5
113 ssd_mobilenet_v1_coco_tf 300x300 2.5 433.5 529.5
114 ssd_mobilenet_v2 360x480 6.6 82.5 157.9
115 ssd_mobilenet_v2_coco_tf 300x300 3.8 387.2 506.7
116 ssd_pedestrian_pruned_0_97 360x360 5.9 383.4 645.7
117 ssd_resnet_50_fpn_coco_tf 640x640 178.4 10.9 12.2
118 ssd_traffic_pruned_0_9 360x480 11.6 337.2 616.7
119 ssdlite_mobilenet_v2_coco_tf 300x300 1.5 385.2 517.7
120 ssr_pt 256x256 39.72 74 77.5
121 superpoint_tf 480x640 52.4 49.8 104.3
122 textmountain_pt 960x960 575.2 20.7 29.6
123 tiny_yolov3_vmss 416x416 5.46 719.7 1424.4
124 tsd_yolox_pt 640x640 73 134 191.1
125 ultrafast_pt 288x800 8.4 392.4 966.1
126 unet_chaos-CT_pt 512x512 23.3 74.6 238.6
127 vehicle_make_resnet18_pt 224x224 3.627 1996.6 5111.8
128 vehicle_type_resnet18_pt 224x224 3.627 2050.7 5155.6
129 vgg_16_tf 224x224 31 531.4 657.2
130 vgg_19_tf 224x224 39.3 482.1 584
131 vpgnet_pruned_0_99 480x640 2.5 357.4 722.6
132 yolov2_voc 448x448 34 418.1 803
133 yolov2_voc_pruned_0_66 448x448 11.6 455.1 1284.1
134 yolov2_voc_pruned_0_71 448x448 9.9 575.4 1330.1
135 yolov2_voc_pruned_0_77 448x448 7.8 598.8 1366
136 yolov3_adas_pruned_0_9 256x512 5.5 524.1 1049.7
137 yolov3_bdd 288x512 53.7 205.3 292.9
138 yolov3_coco_416_tf2 416x416 65.9 179.9 290.1
139 yolov3_voc 416x416 65.4 216.6 289.4
140 yolov3_voc_tf 416x416 65.6 217.6 289.5
141 yolov4_leaky_416_tf 416x416 60.3 141.2 214.6
142 yolov4_leaky_512_tf 512x512 91.2 104.7 154.6
143 yolov4_leaky_spp_m 416x416 60.1 146.1 218.3
144 yolov4_leaky_spp_m_pruned_0_36 416x416 38.2 156.6 241.5