The power consumption of DLA on orin is much higher than that of GPU？

Robotics & Edge Computing Jetson & Embedded Systems Jetson AGX Orin
1

  • Device: nvidia jetson agx orin

  • CNN: Yolov5’s 0-343 layers (all layers support DLA)

  • Method of calculating power consumption: Use the Jetson Power GUI to collect Power Monitor logs on the graphical interface， interval=1s，total power consumption=VDD_ GPU_ SOC, VDD_ CPU_ CV, VIN_ SYS_ 5V0

  • Collection process:

  1. Run yolov5_int8_gpu.engine 1000 times using trtexec, collecting the total power consumption of yolov5 on the GPU (trtexec – loadengine=yolov5_int8_gpu. engine – iterations=1000)
    gpu

  2. Run yolov5_int8_dla.engine 1000 times using trtexec, collecting the total power consumption of yolov5 on the DLA (trtexec – loadengine=yolov5_int8_dla. engine – iterations=1000)

    dla
    dla601×824 50.1 KB

Although the inst power (VDD_GPU_SOC) of the GPU is about 30000mW when running on the GPU, and the inst power (VDD_CPU_CV) of the DLA and CPU is only about 2800mW when running on the DLA, the inference duration of the DLA is much longer than that of the GPU, resulting in a high total power consumption of the DLA.

The final result is to perform 1000 yolov5 inference tests, with a total power consumption of 210027mJ on the GPU and 582913mJ on the DLA, indicating that the DLA does not save energy.

May I ask if there are any errors in my calculation process? If there are no errors in the calculation, why is DLA not as energy-efficient as the official explanation? The accuracy of the model I am using is int8 instead of fp16.

Thanks in advance.
Best regard,
FanLi

3

Hi,

Does your model can run solely on the DLA?
Or does it have some fallback layer that will run on GPU?

Thanks.

4

Hi,

Below is a useful link for your reference.
In our experiment, DLA is about 3~5x power efficient compared to the GPU.

Thanks.

5

Thanks for your prompt reply.
The CNN model I am using is the 0-343 layers of YOLOv5. I can confirm that its converted engine can run completely on DLA and will not fallback to GPU. The following is the result of visualizing yolov5_int8_dla. engine using TensorRT tools.

visual
visual1619×776 69.5 KB

6

This is very useful for me. I will carefully read and learn the content of this link. Thank you again for your help！

9

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.