Any hope getting this to work with Windows Subsystem for Linux?
I got my model up and training on wsl, which really surprised me -- it was super easy and just worked. But I didn't even try venturing into gpu acceleration... perhaps it would "just work" as well? Though I suspect not.
Curiously, I tried and it seems that it couldn't communicate to the GPU. Installation is successful, but upon running `nvidia-smi` command, it returned the following:
NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.
I'll try this later, seems to be targetting pre-Pascal cards (need more info on whether this will work with Cuda 8).
"Cuda for Engineers" (excellent book by Storti/Yurtoglu) uses VS 2013 and the VC++ that comes with it, all my other cuda books (Handbook by Wilt, Wrox Pro Cuda Prog'g) run on linux. Also worth noting that OS X seems to have gotten feature parity with Cuda 8's unified memory (my mac doesn't have recent Nvidia card to test).
I got a 1080 working with VS 2015 (via the CUDA 8 beta). However, I strongly recommend using VS 2013. Frustratingly, CUDA VS 2015 support does not extend to the latest Visual Studio update level. So, to use CUDA I had to forgo the latest versions of useful extensions (e.g. Python Tools) and even forgo MS software updates (e.g. for SQL Studio, which apparently auto-installed the latest VS updates).
This is great. We do a lot of work with GPUs in Win10 at Paperspace (https://www.paperspace.com) but we mostly use GPUs for video acceleration not compute.
That said, all of our cloud VMs have GPUs and we are adding new instance types designed for ML/ HPC.
Reach out directly if you are in the space and want early access to CUDA-backed Windows instances in the cloud.
I got my model up and training on wsl, which really surprised me -- it was super easy and just worked. But I didn't even try venturing into gpu acceleration... perhaps it would "just work" as well? Though I suspect not.