• State TI Thinks Resolved
  • Locked Locked
  • Replies 3 replies
  • Answers 1 answer
  • Subscribers 99 subscribers
  • Views 1564 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Sitara v. iMX8 MobileNet benchmark

Anders Svensson
Intellectual 1155 points
Other Parts Discussed in Thread: AM5729

Hi

I need your help to provide measurements on MobileNet x.x for our Sitara product.

First i need our best competitive device vs IMX8.

Then I need MobileNet x.x figures together with power consumption.

Would be very good to understand the performance and power consumption depending on what part of Sitara that is used to do the MobileNet benchmark (CPU, GPU or DSP)

 

Regards

/Anders

  • 0
    TI__Mastermind 26215 points

    Anders,

    Would you please tell me more about your application? Depending on the application, there are different devices that we can play here.

    What are the most important features from your application perspective? Size? Power Consumption? AI performance?

  • 0 in reply to Rogerio Almeida
    TI__Intellectual 1155 points

    Hi again

    AI performance together with power consumption is everything for this opportunity.

    And we must be better than the IMX 8 device used today.

    I can´t give more information on customer and implementation on this forum.

    Very important . I need the data latest end of July otherwise this is not worth spending time on.

    /Anders

  • 0 in reply to Rogerio Almeida
    TI__Mastermind 22946 points

    Hi Anders,

    AM5729 is the best device in existing Sitara Portfolio to run MobileNet from performance perspective. The device has 2xA15, 2xC66x and 4xEVE (specialized NN accelerator). 

    MobileNet V1 run at 10 fps for 224x224 image classification, when using DSP + EVE alone (No A15 usage). We are optimizing certain layers of MobileNet model on our DSP and EVE, so we expect to some performance improvement in our next Processor SDK release. The model is running in 8 bit fixed pt quantization.

    Regarding power consumption, fully loaded EVE and DSP consumes around 0.5W, but with dual core A15 running linux stack and other peripherals on, SoC level power consumption is in the range of 5W.

    Dual Core A15 @ 1.5 GHz can as well be utilized to run the model using ArmNN. Following performance bench-marking of Mobilenetv1 and v2 on 2xA15. I don't have power benchmarking for this.

    ARMNN/ACL 19.05 

    AM572x (2xA15 @ 1.5 GHz)

    MobilenetV1 (1.0, 224x224x3)

    5.5-6.9

    MobilenetV2 (1.0, 224x224x3)

    7-8.3

    Regards,

    Manisha