Dear TI,
Is there any analyse or document about "numParamBits" and "numFeatureBits" in TIDL quantization?
Is it recomended using 8bit for runtime ? Why ?
Thank you
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.
Dear TI,
Is there any analyse or document about "numParamBits" and "numFeatureBits" in TIDL quantization?
Is it recomended using 8bit for runtime ? Why ?
Thank you
Hi,
numParamBits and numFeatureBits control how many bits are used for quantization for parameters and features/activation respectively. It is important to note that if you make any of them greater than 8 bit the other one will automatically become 16 bit.
Is it recomended using 8bit for runtime ? Why ?
Yes 8 bit quantization is recommended and the reason is because you will get the best performance for 8 bit. For example TDA4VM can give 8 TOPS for 8 bit where as it becomes 2 TOPS for 16 bit. If you are observing any issues with 8 bit quantization then please follow the steps in the below documentation to see how it can be improved :
Regards,
Anshu