Open DHOFM opened 7 years ago
We are not the Azure team so we can't answer your question regarding V100 availability.
We are trying to give 16bit support by this iteration, still on track although with some issues with our build environment.
Any updates on support for V100 (and new Titan V)? Besides the 16bit support, are other changes required to take advantage of the tensor cores?
Please take a look at a WIP branch kedeng/CUDA9 if you want to try on V100. There are some cuda code changes needed to make use of TensorCore.
As mentinoned here: https://github.com/Microsoft/CNTK/issues/2194 you are planning to support Volta GPU. Since CUDA 9 is out, are there Updates ? Will CNTK make use of the 120 TFLOPS (theoretically) ? Will V100 GPUs come to Azure ( they are available in AWS now ) ???
Thanks, Dirk