The different computing reuslts in the same nv-docker container?

Hi everyone,

I have more than one local machine with nv-docker. Now in nv-docker container, I use tensorflow and torch library for accelerating. The question is different local machines with running on the same nv-docker image have the same computing result?

The way I did is

  1. Try keeping the all random seed the same (ex tensorflow.set_random_seed(1234))
  2. Make sure they run on the same docker image
  3. nvidia-driver are the same version and only a gpu is visible (ex os.environ[“CUDA_VISIBLE_DEVICES”] = “0”)

with no luck, I still got the different result. Could someone explain me about? I have got confused a long time.

docker image environment:
ubuntu:16.04
cuda:9.0
cudnn:7.0
tensorflow-gpu=1.12.0
torch=1.1.0

This may be of interest:

https://developer.nvidia.com/gtc/2019/video/S9911