• State Resolved
  • Locked Locked
  • Replies 8 replies
  • Subscribers 100 subscribers
  • Views 352 views
  • Users 0 members are here
Support feedback
Options
Options

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.

TDA4VH-Q1: The inference time of model is different using edge-AI tools and edge-AI studio

Liu Tsing-Yue
Intellectual 291 points
Part Number: TDA4VH-Q1
Other Parts Discussed in Thread: TDA4VH, AM69A

Hi experts,

We analyze our model using edge-AI tools on a TDA4VH (32TOPS), the time is about 7.5ms

We also use model analyzer from edge-AI studio, connected to AM69A (32 TOPS)

dev.ti.com/.../

However, the result has a lot of difference, the inference only needs 2.40ms

What might be the cause of this difference?

For edge-AI tools, the sdk version is 09_00_00_06, and I heard that the sdk version of edge-AI studio is 8.06

Thanks.

  • 0
    TI__Mastermind 24041 points

    Hi,

    Liu Tsing-Yue said:
    For edge-AI tools, the sdk version is 09_00_00_06, and I heard that the sdk version of edge-AI studio is 8.06

    Please make sure that reference point for comparison is correct (SDK version)

    I believe there should be a way to select sdk version on edgeai studio tool.

    Thanks 

  • 0 in reply to Pratik Kedar
    Intellectual 291 points

    Hi,

    I asked this earlier, but unfortunately only sdk 8.6 is supported.

    https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1354406/edge-ai-studio-choosing-sdk-version-in-model-analyzer/5168086?tisearch=e2e-sitesearch&keymatch=%2520user%253A574787#5168086

    We assumed that the inference time will be close, in order to test our model on differenent SOC to get inference time that can be referenced.

    But now the inference time might not be able to referenced.

  • 0 in reply to Liu Tsing-Yue
    TI__Mastermind 24041 points

    Hi,

    Thanks for confirmation, yes seems like tool is on supporting 9.0 sdk version yet.

    Liu Tsing-Yue said:
    We assumed that the inference time will be close,

    No, this is not the case, as every sdk release has updated c7x firmware and other related sdk components.

    I hope the observation related to nuances between benchmark data is clear, as it was rooted to sdk version mismatch.

    Is there anything else you need help with ?

    Thanks 

  • 0 in reply to Pratik Kedar
    Intellectual 291 points

    Hi,

    but the benchmark difference is large,  for the same model, one inference time is 7.5ms and another is 2.4ms.

    Is the influence of sdk version really ths big?

  • 0 in reply to Liu Tsing-Yue
    TI__Mastermind 24041 points

    Hi,

    Could you please share how are benchmarking the model on target ? which flow are you using here ? openvx based, gst based, or TIDL-RT via RTOS sdk ?

    In case if you are using TIDL-RT can you share infer config to look for ?

  • 0 in reply to Pratik Kedar
    Intellectual 291 points

    Hi,

    We both use OSRT for model inference

    https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_osr_debug.md

    here's the config we use:

  • +1 in reply to Liu Tsing-Yue
    TI__Mastermind 24041 points

    Hi,

    Comparison would make sense if we do comparison with same sdk ! as its hard to catch the nuances in settings of both the flow and underneath sdk components are different.

    Still am suspecting this could be because of few reasons (i dont have full visibility of both the flows at my end) 

    1. Different quantization style 

    1.1 8/16 bits, mixed etc 

    1.2 Calibration images set, calibration iterations, quantization style (symmetric vs asymmetric), other aspects of quantization

    2. Arm offload 

    2.1 Check how many nodes are delegated to arm and c7x moreover you can check the subgraphs ! 

    2.1 Are you opting for meta arch support 

    3. Advanced Options 

    Please make sure of the above mentioned things along with remaining compilations options are same.

    Did you tried running same model in 8.5 OSRT flow ? what was the benchmark in that case ? 

  • 0 in reply to Pratik Kedar
    Intellectual 291 points

    Hi,

    I think the above settings are the same.

    However we might not change our current sdk version, thus the testing will stop here.

    Thanks for your support.