Alveo U50/U50LV Data Accelerator Card

The Xilinx® Alveo U50 Data Center accelerator cards are peripheral component interconnect express (PCIe®) Gen3x16 compliant and Gen4x8 compatible cards featuring the Xilinx 16 nm UltraScale+ technology. In this release, DPU is implemented in program logic for deep learning inference acceleration.

Note: Some models cannot run at the highest frequency of DPU and need DPU frequency reduction. See For Edge for DPU frequency reduction operation.

U50 Performance with 6E300 MHz DPUCAHX8H

Refer to the following table for the throughput performance (in frames/sec or fps) for various neural network samples on U50 Gen3x4 with DPUCAHX8H running at 6E@300 MHz.

Table 1. U50 Performance with 6E300 MHz DPUCAHX8H
No	Neural Network	Input Size	GOPS	DPU Frequency (MHz)	Performance (fps) (Multiple thread)
1	densebox_320_320	320x320	0.49	300	2231.2
2	densebox_640_360	360x640	1.1	300	969.4
3	ENet_cityscapes_pt	512x1024	8.6	300x0.9	69.5
4	face_landmark	96x72	0.14	300	10035.1
5	face-quality	80x60	0.06	300x0.9	21786.1
6	face-quality_pt	80x60	0.06	300x0.9	21835.0
7	facerec_resnet20	112x96	3.5	300	1330.5
8	facerec-resnet20_mixed_pt	112x96	3.5	300x0.9	1213.5
9	facerec_resnet64	112x96	11	300	491.1
10	facereid-large_pt	96x96	0.5	300x0.9	7096.1
11	facereid-small_pt	80x80	0.09	300x0.9	19550.9
12	fpn	256x512	8.9	300	439.2
13	FPN_Res18_Medical_segmentation	320x320	45.3	300	101.5
14	FPN-resnet18_covid19-seg_pt	352x352	22.7	300x0.9	205.1
15	inception_resnet_v2_tf	299x299	26.4	300	172.1
16	inception_v1	224x224	3.2	300	1232.8
17	inception_v1_tf	224x224	3	300	1252.0
18	inception_v2	224x224	4	300	972.4
19	inception_v3	299x299	11.4	300	406.5
20	inception_v3_pt	299x299	5.7	300	406.6
21	inception_v3_tf	299x299	11.5	300	406.9
22	inception_v3_tf2	299x299	11.5	300x0.9	372.0
23	inception_v4	299x299	24.5	300	185.0
24	inception_v4_2016_09_09_tf	299x299	24.6	300	184.8
25	medical_seg_cell_tf2	128x128	5.3	300x0.9	1051.1
26	MLPerf_resnet50_v1.5_tf	224x224	8.19	300x0.9	500.2
27	mlperf_ssd_resnet34_tf	1200x1200	433	300x0.9	13.8
28	multi_task	288x512	14.8	300	336.4
29	openpose_pruned_0_3	368x368	49.9	300x0.9	29.1
30	personreid-res18_pt	176x80	1.1	300x0.9	3404.4
31	personreid-res50_pt	256x128	5.4	300x0.9	805.8
32	plate_detection	320x320	0.49	300	5701.5
33	plate_num	96x288	1.75	300x0.9	1184.2
34	pmg_pt	224x224	2.28	300	1090.0
35	refinedet_baseline	480x360	123	300x0.9	49.8
36	RefineDet-Medical_EDD_tf	320x320	9.8	300x0.9	421.0
37	refinedet_pruned_0_8	360x480	25	300x0.9	193.0
38	refinedet_pruned_0_92	360x480	10.1	300x0.9	404.4
39	refinedet_pruned_0_96	360x480	5.1	300x0.9	592.6
40	refinedet_VOC_tf	320x320	81.9	300x0.9	71.2
41	reid	80x160	0.95	300	3849.5
42	resnet18	224x224	3.7	300	1410.3
43	resnet50	224x224	7.7	300	572.7
44	resnet50_pt	224x224	4.1	300	554.7
45	resnet50_tf2	224x224	7.7	300x0.9	516.4
46	resnet_v1_101_tf	224x224	14.4	300	333.6
47	resnet_v1_152_tf	224x224	21.8	300	222.3
48	resnet_v1_50_tf	224x224	7	300	642.6
49	salsanext_pt	64x2048	20.4	300x0.9	143.6
50	salsanext_v2_pt	64x2048	32	300	25.1
51	SemanticFPN_cityscapes_pt	256x512	10	300x0.9	425.8
52	semantic_seg_citys_tf2	512x1024	54	300x0.9	50.0
53	sp_net	128x224	0.55	300	2826.2
54	squeezenet	227x227	0.76	300	3619.4
55	squeezenet_pt	224x224	0.82	300	2079.3
56	ssd_adas_pruned_0_95	360x480	6.3	300	654.5
57	ssd_pedestrian_pruned_0_97	360x360	5.9	300x0.9	568.6
58	ssd_resnet_50_fpn_coco_tf	640x640	178.4	300x0.9	31.2
59	ssd_traffic_pruned_0_9	360x480	11.6	300	428.9
60	tiny_yolov3_vmss	416x416	5.46	300x0.9	870.6
61	unet_chaos-CT_pt	512x512	23.3	300x0.9	60.2
62	vgg_16_tf	224x224	31	300	156.3
63	vgg_19_tf	224x224	39.3	300	131.0
64	vpgnet_pruned_0_99	480x640	2.5	300x0.9	480.3
65	yolov2_voc	448x448	34	300x0.9	162.4
66	yolov2_voc_pruned_0_66	448x448	11.6	300x0.9	411.3
67	yolov2_voc_pruned_0_71	448x448	9.9	300x0.9	481.3
68	yolov2_voc_pruned_0_77	448x448	7.8	300x0.9	585.4
69	yolov3_adas_pruned_0_9	256x512	5.5	300x0.9	611.5
70	yolov3_bdd	288x512	53.7	300x0.9	74.9
71	yolov3_voc	416x416	65.4	300x0.9	76.7
72	yolov3_voc_tf	416x416	65.6	300x0.9	77.2
73	yolov4_leaky_spp_m	416x416	60.1	300x0.9	80.8
74	yolov4_leaky_spp_m_pruned_0_36	416x416	38.2	300	84.8

U50 Performance with 1E333 MHz DPUCAHX8L

Refer to the following table for the throughput performance (in frames/sec or fps) for various neural network samples on U50 Gen3x4 with DPUCAHX8L running at 1E@333 MHz.

Table 2. U50 Performance with 1E333 MHz DPUCAHX8L
No	Neural Network	Input Size	GOPS	DPU Frequency (MHz)	Performance (fps) (Multiple thread)
1	ENet_cityscapes_pt	512x1024	8.6	333	5.9
2	face_landmark	96x72	0.14	333	4988.1
3	face-quality	80x60	0.06	333	7036.6
4	face-quality_pt	80x60	0.06	333	7037.8
5	facerec_resnet20	112x96	3.5	333	323.0
6	facerec-resnet20_mixed_pt	112x96	3.5	333	321.7
7	facerec_resnet64	112x96	11	333	144.3
8	facereid-small_pt	80x80	0.09	333	4403.3
9	fpn	256x512	8.9	333	39.8
10	FPN_Res18_Medical_segmentation	320x320	45.3	333	14.9
11	FPN-resnet18_covid19-seg_pt	352x352	22.7	333	90.3
12	inception_resnet_v2_tf	299x299	26.4	333	35.2
13	inception_v1	224x224	3.2	333	343.5
14	inception_v1_tf	224x224	3	333	348.3
15	inception_v2	224x224	3.88	333	197.3
16	inception_v3	299x299	11.4	333	107.5
17	inception_v3_pt	299x299	5.7	333	107.2
18	inception_v3_tf	299x299	11.5	333	107.8
19	inception_v3_tf2	299x299	11.5	333	102.1
20	inception_v4	299x299	24.5	333	57.0
21	inception_v4_2016_09_09_tf	299x299	24.6	333	57.1
22	medical_seg_cell_tf2	128x128	5.3	333	116.6
23	MLPerf_resnet50_v1.5_tf	224x224	8.19	333	88.9
24	mlperf_ssd_resnet34_tf	1200x1200	433	333	7.0
25	mobilenet_1_0_224_tf2	224x224	1.1	333	2104.1
26	mobilenet_v1_0_5_160_tf	160x160	0.15	333	5480.7
27	mobilenet_v1_1_0_224_tf	224x224	1.1	333	2135.2
28	mobilenet_v2	224x224	0.6	333	1190.2
29	mobilenet_v2_1_0_224_tf	224x224	0.6	333	1186.5
30	mobilenet_v2_1_4_224_tf	224x224	1.2	333	872.4
31	multi_task	288x512	14.8	333	21.5
32	openpose_pruned_0_3	368x368	49.9	333	17.4
33	personreid-res50_pt	256x128	5.4	333	106.4
34	plate_detection	320x320	0.49	333	1036.7
35	rcan_pruned_tf	360x640	86.95	333	4.1
36	refinedet_baseline	480x360	123	333	30.1
37	RefineDet-Medical_EDD_tf	320x320	9.8	333	152.3
38	refinedet_pruned_0_8	360x480	25	333	67.4
39	refinedet_pruned_0_92	360x480	10.1	333	77.4
40	refinedet_pruned_0_96	360x480	5.1	333	94.5
41	refinedet_VOC_tf	320x320	81.9	333	48.0
42	reid	80x160	0.95	333	612.4
43	resnet18	224x224	3.7	333	320.9
44	resnet50	224x224	7.7	333	87.9
45	resnet50_pt	224x224	4.1	333	89.2
46	resnet50_tf2	224x224	7.7	333	88.0
47	resnet_v1_101_tf	224x224	14.4	333	57.8
48	resnet_v1_152_tf	224x224	21.8	333	38.6
49	resnet_v1_50_tf	224x224	7	333	101.7
50	retinaface	360x640	1.11	333	206.5
51	salsanext_pt	64x2048	20.4	333	12.0
52	SemanticFPN_cityscapes_pt	256x512	10	333	36.5
53	SemanticFPN_Mobilenetv2_pt	512x1024	5.4	333	11.1
54	semantic_seg_citys_tf2	512x1024	54	333	5.6
55	sp_net	128x224	0.55	333	10455.9
56	squeezenet	227x227	0.76	333	838.0
57	squeezenet_pt	224x224	0.82	333	471.2
58	ssd_adas_pruned_0_95	360x480	6.3	333	117.5
59	ssdlite_mobilenet_v2_coco_tf	300x300	1.5	333	610.7
60	ssd_mobilenet_v1_coco_tf	300x300	2.5	333	1053.1
61	ssd_mobilenet_v2	360x480	6.6	333	157.7
62	ssd_mobilenet_v2_coco_tf	300x300	3.8	333	238.4
63	ssd_pedestrian_pruned_0_97	360x360	5.9	333	107.9
64	ssd_traffic_pruned_0_9	360x480	11.6	333	112.5
65	vgg_16_tf	224x224	31	333	57.7
66	vgg_19_tf	224x224	39.3	333	51.9
67	vpgnet_pruned_0_99	480x640	2.5	333	72.3

U50LV Performance with 10E275 MHz DPUCAHX8H

The following table shows the throughput performance (in frames/sec or fps) for various neural network samples on U50LV Gen3x4 with DPUCAHX8H running at 10E@275 MHz.

Table 3. U50LV Performance with 10E275 MHz DPUCAHX8H
No	Neural Network	Input Size	GOPS	DPU Frequency (MHz)	Performance (fps) (Multiple thread)
1	densebox_320_320	320x320	0.49	275	3247.4
2	densebox_640_360	360x640	1.1	275	1435.5
3	ENet_cityscapes_pt	512x1024	8.6	275x0.9	104.1
4	face_landmark	96x72	0.14	275	14849.1
5	face-quality	80x60	0.06	275x0.9	23500.4
6	face-quality_pt	80x60	0.06	275x0.9	23449.1
7	facerec_resnet20	112x96	3.5	275	2026.2
8	facerec-resnet20_mixed_pt	112x96	3.5	275x0.9	1847.0
9	facerec_resnet64	112x96	11	275	749.7
10	facereid-large_pt	96x96	0.5	275x0.9	10587.9
11	facereid-small_pt	80x80	0.09	275x0.9	25648.8
12	fpn	256x512	8.9	275x0.9	597.8
13	FPN_Res18_Medical_segmentation	320x320	45.3	275x0.9	140.0
14	FPN-resnet18_covid19-seg_pt	352x352	22.7	275x0.9	313.2
15	inception_resnet_v2_tf	299x299	26.4	275x0.8	210.7
16	inception_v1	224x224	3.2	275x0.9	1700.5
17	inception_v1_tf	224x224	3	275x0.9	1726.9
18	inception_v2	224x224	4	275x0.9	1342.4
19	inception_v3	299x299	11.4	275x0.8	498.2
20	inception_v3_pt	299x299	5.7	275x0.8	498.0
21	inception_v3_tf	299x299	11.5	275x0.8	498.4
22	inception_v3_tf2	299x299	11.5	275x0.9	567.6
23	inception_v4	299x299	24.5	275x0.8	226.7
24	inception_v4_2016_09_09_tf	299x299	24.6	275x0.8	226.7
25	medical_seg_cell_tf2	128x128	5.3	275x0.9	1600.3
26	MLPerf_resnet50_v1.5_tf	224x224	8.19	275x0.9	763.1
27	mlperf_ssd_resnet34_tf	1200x1200	433	275x0.9	21.1
28	multi_task	288x512	14.8	275x0.9	462.2
29	openpose_pruned_0_3	368x368	49.9	275x0.7	34.8
30	personreid-res18_pt	176x80	1.1	275x0.9	5144.3
31	personreid-res50_pt	256x128	5.4	275x0.9	1227.6
32	plate_detection	320x320	0.49	275	6997.9
33	plate_num	96x288	1.75	275x0.9	1656.7
34	pmg_pt	224x224	2.28	275x0.8	1333.7
35	refinedet_baseline	480x360	123	275x0.7	59.2
36	RefineDet-Medical_EDD_tf	320x320	9.8	275x0.9	641.2
37	refinedet_pruned_0_8	360x480	25	275x0.9	295.6
38	refinedet_pruned_0_92	360x480	10.1	275x0.9	614.2
39	refinedet_pruned_0_96	360x480	5.1	275x0.9	902.6
40	refinedet_VOC_tf	320x320	81.9	275x0.9	108.6
41	reid	80x160	0.95	275	5833.1
42	resnet18	224x224	3.7	275x0.9	1935.5
43	resnet50	224x224	7.7	275x0.9	787.5
44	resnet50_pt	224x224	4.1	275x0.9	763.6
45	resnet50_tf2	224x224	7.7	275x0.9	788.3
46	resnet_v1_101_tf	224x224	14.4	275x0.9	459.7
47	resnet_v1_152_tf	224x224	21.8	275x0.9	306.5
48	resnet_v1_50_tf	224x224	7	275x0.9	884.5
49	salsanext_pt	64x2048	20.4	275x0.9	149.7
50	salsanext_v2_pt	64x2048	32	275x0.8	35.9
51	SemanticFPN_cityscapes_pt	256x512	10	275x0.9	640.8
52	semantic_seg_citys_tf2	512x1024	54	275x0.9	76.5
53	sp_net	128x224	0.55	275	4291.3
54	squeezenet	227x227	0.76	275x0.9	4826.3
55	squeezenet_pt	224x224	0.82	275x0.9	2847.2
56	ssd_adas_pruned_0_95	360x480	6.3	275x0.8	821.0
57	ssd_pedestrian_pruned_0_97	360x360	5.9	275x0.9	863.6
58	ssd_resnet_50_fpn_coco_tf	640x640	178.4	275x0.8	42.3
59	ssd_traffic_pruned_0_9	360x480	11.6	275	653.4
60	tiny_yolov3_vmss	416x416	5.46	275x0.9	1324.6
61	unet_chaos-CT_pt	512x512	23.3	275x0.9	91.4
62	vgg_16_tf	224x224	31	275x0.9	218.6
63	vgg_19_tf	224x224	39.3	275x0.9	182.7
64	vpgnet_pruned_0_99	480x640	2.5	275x0.9	720.1
65	yolov2_voc	448x448	34	275x0.8	220.8
66	yolov2_voc_pruned_0_66	448x448	11.6	275x0.8	558.5
67	yolov2_voc_pruned_0_71	448x448	9.9	275x0.8	654.1
68	yolov2_voc_pruned_0_77	448x448	7.8	275x0.8	794.0
69	yolov3_adas_pruned_0_9	256x512	5.5	275x0.8	835.4
70	yolov3_bdd	288x512	53.7	275x0.8	101.8
71	yolov3_voc	416x416	65.4	275x0.8	104.4
72	yolov3_voc_tf	416x416	65.6	275x0.8	104.7
73	yolov4_leaky_spp_m	416x416	60.1	275x0.8	110.0
74	yolov4_leaky_spp_m_pruned_0_36	416x416	38.2	275x0.8	112.2

U50LV Performance with 1E250 MHz DPUCAHX8L

The following table shows the throughput performance (in frames/sec or fps) for various neural network samples on U50LV Gen3x4 with DPUCAHX8L running at 1E@250 MHz.

Table 4. U50LV Performance with 1E250 MHz DPUCAHX8L
No	Neural Network	Input Size	GOPS	DPU Frequency (MHz)	Performance (fps) (Multiple thread)
1	ENet_cityscapes_pt	512x1024	8.6	250	5.7
2	face_landmark	96x72	0.14	250	4592.4
3	face-quality	80x60	0.06	250	6584.9
4	face-quality_pt	80x60	0.06	250	6538.5
5	facerec_resnet20	112x96	3.5	250	295.3
6	facerec-resnet20_mixed_pt	112x96	3.5	250	293.8
7	facerec_resnet64	112x96	11	250	129.0
8	facereid-small_pt	80x80	0.09	250	4237.6
9	fpn	256x512	8.9	250	38.9
10	FPN_Res18_Medical_segmentation	320x320	45.3	250	14.2
11	FPN-resnet18_covid19-seg_pt	352x352	22.7	250	82.1
12	inception_resnet_v2_tf	299x299	26.4	250	32.2
13	inception_v1	224x224	3.2	250	322.5
14	inception_v1_tf	224x224	3	250	326.9
15	inception_v2	224x224	3.88	250	185.0
16	inception_v3	299x299	11.4	250	100.0
17	inception_v3_pt	299x299	5.7	250	99.7
18	inception_v3_tf	299x299	11.5	250	99.9
19	inception_v3_tf2	299x299	11.5	250	93.9
20	inception_v4	299x299	24.5	250	52.7
21	inception_v4_2016_09_09_tf	299x299	24.6	250	52.6
22	medical_seg_cell_tf2	128x128	5.3	250	114.0
23	MLPerf_resnet50_v1.5_tf	224x224	8.19	250	80.9
24	mlperf_ssd_resnet34_tf	1200x1200	433	250	6.2
25	mobilenet_1_0_224_tf2	224x224	1.1	250	1678.8
26	mobilenet_v1_0_5_160_tf	160x160	0.15	250	4660.9
27	mobilenet_v1_1_0_224_tf	224x224	1.1	250	1714.1
28	mobilenet_v2	224x224	0.6	250	1012.4
29	mobilenet_v2_1_0_224_tf	224x224	0.6	250	1008.8
30	mobilenet_v2_1_4_224_tf	224x224	1.2	250	742.2
31	multi_task	288x512	14.8	250	21.1
32	openpose_pruned_0_3	368x368	49.9	250	14.4
33	personreid-res50_pt	256x128	5.4	250	99.1
34	plate_detection	320x320	0.49	250	1104.5
35	rcan_pruned_tf	360x640	86.95	250	4.0
36	refinedet_baseline	480x360	123	250	26.3
37	RefineDet-Medical_EDD_tf	320x320	9.8	250	145.2
38	refinedet_pruned_0_8	360x480	25	250	64.6
39	refinedet_pruned_0_92	360x480	10.1	250	76.2
40	refinedet_pruned_0_96	360x480	5.1	250	93.7
41	refinedet_VOC_tf	320x320	81.9	250	39.8
42	reid	80x160	0.95	250	569.5
43	resnet18	224x224	3.7	250	296.6
44	resnet50	224x224	7.7	250	80.5
45	resnet50_pt	224x224	4.1	250	81.2
46	resnet50_tf2	224x224	7.7	250	80.4
47	resnet_v1_101_tf	224x224	14.4	250	52.4
48	resnet_v1_152_tf	224x224	21.8	250	34.9
49	resnet_v1_50_tf	224x224	7	250	93.3
50	retinaface	360x640	1.11	250	197.1
51	salsanext_pt	64x2048	20.4	250	11.6
52	SemanticFPN_cityscapes_pt	256x512	10	250	35.6
53	SemanticFPN_Mobilenetv2_pt	512x1024	5.4	250	11.0
54	semantic_seg_citys_tf2	512x1024	54	250	5.4
55	sp_net	128x224	0.55	250	10520.8
56	squeezenet	227x227	0.76	250	793.6
57	squeezenet_pt	224x224	0.82	250	471.6
58	ssd_adas_pruned_0_95	360x480	6.3	250	118.3
59	ssdlite_mobilenet_v2_coco_tf	300x300	1.5	250	522.5
60	ssd_mobilenet_v1_coco_tf	300x300	2.5	250	854.4
61	ssd_mobilenet_v2	360x480	6.6	250	138.0
62	ssd_mobilenet_v2_coco_tf	300x300	3.8	250	210.1
63	ssd_pedestrian_pruned_0_97	360x360	5.9	250	109.0
64	ssd_traffic_pruned_0_9	360x480	11.6	250	111.3
65	vgg_16_tf	224x224	31	250	52.2
66	vgg_19_tf	224x224	39.3	250	46.1
67	vpgnet_pruned_0_99	480x640	2.5	250	74.0