Does this mitigate the requirement of installing CuDNN?

microsoft / tensorflow-directml

Fork of TensorFlow accelerated by DirectML

Apache License 2.0

456 stars 32 forks source link

Does this mitigate the requirement of installing CuDNN? #17

Closed roychowdhuryrohit-dev closed 4 years ago

roychowdhuryrohit-dev commented 4 years ago

If I use DirectML, do I still need to install the CUDA libraries? What about support for upcoming RDNA2 AMD GPUs?

jstoecker commented 4 years ago

There is no need to install CUDA or cuDNN when using DirectML.

All DirectX 12-capable GPUs support DirectML, and that definitely includes the upcoming RDNA2 AMD GPUs.

roychowdhuryrohit-dev commented 4 years ago

That’s a pretty big news, thanks for the platform-agnostic support. This solves lot of existing problems that have been prevalent in local AI training ecosystem. I hope this is also true for WSL atleast.

jstoecker commented 4 years ago

Cross-vendor support is fundamental to the DirectX platform as a whole, and that includes both Windows and WSL! We have preview drivers for AMD, Intel, and NVIDIA (see this page).

roychowdhuryrohit-dev commented 4 years ago

I hope everyone(developers, researchers, students) start migrating to DirectML soon. 👍