How to use Tensorflow with AMD Vega 10 GPU

Question

Hello,

I have been using Azure virtual machines to train deep learning models in Tensorflow. Recently, I created an NVas v4 machine with the Data Science VM Ubuntu image and am unable to use the GPU. My understanding is that the NVas v4 GPU is an AMD Vega 10 (as indicated by the output of lshw -class display). Therefore, because the GPU is not made by NVIDIA the usual drivers and libraries that Tensorflow would use to recognize the GPU (CUDA, cuDNN) are irrelevant. Are there additional packages/libraries needed to use Tensorflow with an AMD GPU? It seems odd that Microsoft would offer VMs with GPUs that do not have the necessary software to be utilized.

Mike

Answer 1

To use Tensorflow on AMD GPUs, you can use ROCm.

https://github.com/RadeonOpenCompute/ROCm