Identical behaviour on CPU/GPU with integerization

Hi,

I’m working on integerized implementation of a neural network. Goal is to have identical behavior across platforms and also for CPU/GPU. Which to the best of my knowledge can only be done with integer operations. This is only needed for inference.

I’ve been studying the cudNN API, especially API Reference :: NVIDIA Deep Learning cuDNN Documentation

I believe the only way to do what I want is to use the INT8_CONFIG. This means GPU should support DP4A. Am I correct so far? Or are there other issues that will stand in the way of integer only operations on GPU?

Where can I find list of GPUs that support DP4A? My current GPU is a TITAN RTX, I believe it does not support this.

Thanks in advance!

Hi,

Please refer support matrix document which may help you.

Yes.

Thank you.

Hi,

I had a look at the matrix document. It does not help directly, but it has a link to here: CUDA GPUs | NVIDIA Developer
From this I gather that my Titan RTX has compute capability 7.5 which should include DP4A.

Thanks!