This is the latest (main) BeagleBoard documentation. If you are looking for stable releases, use the drop-down menu on the bottom-left and select the desired version.

Datasheet

This chapter describes the performance measurements of the Edge AI Inference demos.

Performance data of the demos can be auto generated by running following command on target:

debian@beaglebone:/opt/edge_ai_apps/tests# ./gen_data_sheet.sh

The performance measurements includes the following

  1. FPS : Effective framerate at which the application runs

  2. Total time : Average time taken to process each frame, which includes pre-processing, inference and post-processing time

  3. Inference time : Average time taken to infer each frame

  4. CPU loading : Loading on different CPU cores present

  5. DDR BW : DDR read and write BW used

  6. HWA Loading : Loading on different Hardware accelerators present

Following are the latest performance numbers of the C++ demos:

Source : USB Camera

Capture Framerate : 30 fps Resolution : 720p format : JPEG

../../../../_images/edgeai_object_detection.png

Fig. 92 GStreamer based data-flow pipeline with USB camera input and display output

Model

FPS

Total time (ms)

Inference time (ms)

A72 Load (%)

DDR Read BW (MB/s)

DDR Write BW (MB/s)

DDR Total BW (MB/s)

C71 Load (%)

C66_1 Load (%)

C66_2 Load (%)

MCU2_0 Load (%)

MCU2_1 Load (%)

MSC_0 (%)

MSC_1 (%)

VISS (%)

NF (%)

LDC (%)

SDE (%)

DOF (%)

ONR-CL-6150-mobileNetV2-1p4-qat

30.80

33.22

3.02

21.60

1596

619

2215

9.0

20.0

9.0

6.0

1.0

22.17

0

0

0

0

0

0

TFL-CL-0000-mobileNetV1-mlperf

30.69

33.19

1.04

15.93

1425

563

1988

5.0

22.0

9.0

6.0

1.0

21.90

0

0

0

0

0

0

TFL-OD-2020-ssdLite-mobDet-DSP-coco-320x320

30.69

33.25

5.00

10.24

1534

570

2104

15.0

29.0

9.0

6.0

1.0

22.67

0

0

0

0

0

0

TVM-CL-3410-gluoncv-mxnet-mobv2

30.58

33.21

2.02

22.80

1522

617

2139

6.0

20.0

9.0

6.0

1.0

21.84

0

0

0

0

0

0

Source : Video

Video Framerate : 30 fps Resolution : 720p Encoding : h264

../../../../_images/edgeai_video_source.png

Fig. 93 GStreamer based data-flow pipeline with video file input source and display output

Model

FPS

Total time (ms)

Inference time (ms)

A72 Load (%)

DDR Read BW (MB/s)

DDR Write BW (MB/s)

DDR Total BW (MB/s)

C71 Load (%)

C66_1 Load (%)

C66_2 Load (%)

MCU2_0 Load (%)

MCU2_1 Load (%)

MSC_0 (%)

MSC_1 (%)

VISS (%)

NF (%)

LDC (%)

SDE (%)

DOF (%)

ONR-CL-6150-mobileNetV2-1p4-qat

30.52

33.46

3.03

14.28

990

403

1393

2.0

7.0

4.0

1.0

1.0

10.27

0

0

0

0

0

0

TFL-CL-0000-mobileNetV1-mlperf

30.77

33.47

1.07

30.76

746

97

843

2.0

2.0

1.0

1.0

1.0

15.76

0

0

0

0

0

0

TFL-OD-2020-ssdLite-mobDet-DSP-coco-320x320

30.56

33.54

5.06

22.58

736

92

828

2.0

2.0

1.0

1.0

1.0

16.9

0

0

0

0

0

0

TVM-CL-3410-gluoncv-mxnet-mobv2

30.64

33.47

2.01

33.33

712

110

822

1.0

1.0

0.0

1.0

1.0

15.3

0

0

0

0

0

0

Source : CSI Camera (ov5640)

Capture Framerate : 30 fps Resolution : 720p format : YUYV

../../../../_images/edgeai_ov5640_camera_source.png

Fig. 94 GStreamer based data-flow pipeline for with CSI camera (OV5640) input and display output

Model

FPS

Total time (ms)

Inference time (ms)

A72 Load (%)

DDR Read BW (MB/s)

DDR Write BW (MB/s)

DDR Total BW (MB/s)

C71 Load (%)

C66_1 Load (%)

C66_2 Load (%)

MCU2_0 Load (%)

MCU2_1 Load (%)

MSC_0 (%)

MSC_1 (%)

VISS (%)

NF (%)

LDC (%)

SDE (%)

DOF (%)

ONR-CL-6150-mobileNetV2-1p4-qat

29.57

34.09

3.02

12.21

1671

699

2370

8.0

45.0

9.0

6.0

1.0

21.35

0

0

0

0

0

0

TFL-CL-0000-mobileNetV1-mlperf

29.41

34.15

1.01

10.27

1502

645

2147

5.0

47.0

9.0

6.0

1.0

20.96

0

0

0

0

0

0

TFL-OD-2020-ssdLite-mobDet-DSP-coco-320x320

29.36

34.65

5.00

10.5

1610

655

2265

14.0

53.0

9.0

6.0

1.0

21.47

0

0

0

0

0

0

TVM-CL-3410-gluoncv-mxnet-mobv2

29.38

34.17

2.01

11.66

1596

698

2294

6.0

45.0

9.0

5.0

1.0

21.10

0

0

0

0

0

0

Source : CSI Camera with VISS (imx219)

Capture Framerate : 30 fps Resolution : 1080p format : SRGGB8

../../../../_images/edgeai_rpi_camera_source.png

Fig. 95 GStreamer based data-flow pipeline with IMX219 sensor, ISP and display

Model

FPS

Total time (ms)

Inference time (ms)

A72 Load (%)

DDR Read BW (MB/s)

DDR Write BW (MB/s)

DDR Total BW (MB/s)

C71 Load (%)

C66_1 Load (%)

C66_2 Load (%)

MCU2_0 Load (%)

MCU2_1 Load (%)

MSC_0 (%)

MSC_1 (%)

VISS (%)

NF (%)

LDC (%)

SDE (%)

DOF (%)

ONR-CL-6150-mobileNetV2-1p4-qat

30.64

33.19

3.01

15.72

1781

853

2634

9.0

16.0

9.0

13.0

1.0

31.78

0

22.37

0

0

0

0

TFL-CL-0000-mobileNetV1-mlperf

30.59

33.14

1.04

12.78

1612

798

2410

5.0

18.0

9.0

13.0

1.0

31.65

0

22.31

0

0

0

0

TFL-OD-2020-ssdLite-mobDet-DSP-coco-320x320

30.56

33.07

5.00

13.30

1730

809

2539

15.0

25.0

9.0

13.0

1.0

32.6

0

22.19

0

0

0

0

TVM-CL-3410-gluoncv-mxnet-mobv2

30.48

33.14

2.01

12.91

1708

852

2560

7.0

16.0

9.0

13.0

1.0

31.83

0

22.26

0

0

0

0