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.

TDA4VM: TIDL inference: VH performance is lower than VM

Part Number: TDA4VM
Other Parts Discussed in Thread: TDA4VH,

Model: int16-tmp.onnx
1. Model Importer Operation
shell command : ./tidl_model_import.out tidl_import_cfg.txt

2. Model Inference Operation

Environment: TDA4VM and TDA4VH with sdk 8.5.0
Compared with the VM version(j721e-evm-08_05_00_11), Inference performance of the model degrades for VH version.
Execute the test_tidl test case, count the execution time of vxProcessGraph, and find that the performance of the VH version is degraded (VM: 15.571ms; VH: 18.524ms);
Enable C7x L2Cache in TDA4VH, set it to 64KB, re-execute the test case, the execution time is 16.895ms, the performance has improved, but it is still worse than the VM version.

The results are shown in the picture below.

Please refer to the attachment int16-tmp.rar for more details about the configuration, model and test caseint16-tmp.rar

  • Hi,

    The performance of TDA4VM and TDA4VH cannot be directly compared due to following reasons:

    1) TDA4VM has higher MSMC memory compared to TDA4VH individual DSP memory.

    2) We have certain architecture improvements in TDA4VH which help in improving performance compared to TDA4VM.

    Both above points are working in opposite directions and hence final performance seen may depend on SoC architecture and above comparison behaviour may vary from model to model.

    That said, let me note here that we are seeing some performance improvements in SDK 8.6 compared to SDK 8.5 for TDA4VH, so you could see a better performance for this model in 8.6 which will be released shortly.

    Regards,

    Anand