CUDA-python-tensorflow recommended configuration

Hello - we are in the process of standing up a GPU server for DNN model development and scoring and having some headaches with configuration. We have 2 Tesla Volta 100 16GB cards in a Windows Server 2012 R2 OS. We intend on using tensorflow-gpu and keras in a 64 bit python environment. At present we have been using python 3.6.2. Does anyone have any specific configuration information and advice with regards to versions and procedures? We are making the jump from CPU to GPU computing.

One idea: switch to Linux

once you’ve done that, use NGC:

https://docs.nvidia.com/ngc/index.html

Hi Robert - thank you for getting back. Unfortunately I don’t think that will be an option at this time. Its been a couple of years of overcoming multiple hurdles to get to this point, and I’m not sure that architecture would be eager to take on another one-off support target. Do you know of another way within a windows server 12 environment?

We rolled back to CUDA 9.0 from 10.0, and it looks like python tensorflow-gpu package is recognizing the gpu devices. I’m working to finalize installed packages time now and start benchmarking. Looking next at multi-gpu support and some way to monitor GPU utilization

On windows, all of the pre-built TF wheels that I aware of use CUDA 9.0 at this time. So using CUDA 9.0 on your machine is probably a good choice.

I’m not aware of any blockers or issues you would have in using multiple GPUs in that setup.

You can trivially monitor GPU utilization with the nvidia-smi command line tool. It should be installed along with your GPU driver install. If you can’t find it from a command prompt, try searching your drive for nvidia-smi.exe

Once you’ve found the utility, you can use the --help switch to get an idea of some of the features it supports.