TDA4VH-Q1: 16-bit model is 10x slower than 8-bit model

Mu Chen

Part Number: TDA4VH-Q1

Our 16-bit quantized model is approximately 10x slower than 8-bit quantized model. We are using TIDL tools 9.2.6.0.

Below are the complete model conversion logs and model inference logs (on VH, debug_level=2)

0 Pratik Kedar over 1 year ago

Hi,

Adam could you please take initial glance on this issue ?

Lets connect over Webex to discuss this further.

Thanks

Pratik

0 Mu Chen over 1 year ago in reply to Pratik Kedar

Hi,

Is there any update on the issue? or any potential method we can try to improve the model speed?

Thanks

+1 Adam Hua over 1 year ago in reply to Mu Chen

Hi,

As we talked before, 8bit importing is enough for your current application after adding batch norm after conv. So is this problem still valid?

Regards,

Adam

0 Mu Chen over 1 year ago

Hi,

Yes, currently by adding BN we can speedup the 16-bit model to a acceptable speed. This problem can be closed.

Thanks

Processors forum