• State
  • Locked Locked
  • Replies 4 replies
  • Subscribers 99 subscribers
  • Views 187 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.

AM62L: am62lp taking 50% cpu on video encode and decode

Sagar Khimani
Prodigy 20 points

Part Number: AM62L
Other Parts Discussed in Thread: AM62P

Hi,

i'm using ti's am62lp evk and running ai inference on that and i have noticed that video encoder and decoder is utilising 50% cpu. so my ai inference part runing on rest of the cpu and i'm getting less output fps.

can you help me to increase my out fps ?

  • 0
    TI__Mastermind 35970 points

    Hi Sagar,

    On AM62L, we don't have a dedicated video accelerator that can do decode/encode. We rely upon open source software decoders and encoders (ffmpeg libav) which are part of the SDK. Also, since its just a dual core device, you are limited to simple applications like IOT, HMI and general purpose applications.

    Have you tried to look at any other SoC (AM62P/AM62A) for your usecase? Can you elaborate on what sort of application are you trying to use AM62L for?

    Best Regards,

    Suren

  • 0 in reply to Suren Porwar
    Prodigy 20 points

    i'm using am62lp evk and want to run object detection ai model on it.

  • 0 in reply to Sagar Khimani
    TI__Mastermind 35970 points

    Hi Sagar,

    Let me route your query to my expert. Apologies for the delay as the team was out for year end vacation.

    Expect a response soon.

    Best Regards,

    Suren

  • 0 in reply to Suren Porwar
    TI__Mastermind 19716 points

    Hi Sagar,

    Can you clarify if the EVM you are using is [1] or [2]? Those are different processors, but the AM62LP EVM has the AM62x instead of AM62L processor.

    Sagar Khimani said:
    object detection ai model on it.

    For most off-the-shelf, open source models, this is going to be slow. For instance, a yolox-pico at 320x320 resolution is only going to get a few FPS on 2x A53s at 1.4 GHz. For a smooth-looking application, you would need to do some application-level multithreading so that AI can run in a background task at lower rate. This is left to the application developer to implement, but I can point to an example or two for how we have demo'd this with gstreamer on another device (AM62A)

    Some 3rd party software providers can get much higher FPS on Arm cores because they have optimized the model and runtime. Otherwise, our support does not extend to optimizing AI models for CPU execution

    Running software video encoder and decoder will indeed raise CPU utilization -- AM62x and AM62L enable this purely in software. If video codec is critical to your application, then I echo Suren's suggestion to check out AM62A or AM62P. Both have video codec, and the former also has ISP and AI acceleration.  

    BR,
    Reese

    [1] https://www.ti.com/tool/TMDS62LEVM -- Holds the AM62L processor

    [2] https://www.ti.com/tool/SK-AM62-LP -- holds the AM62x processor, but board is 'low power'