Affordable and Optimal Hardware for Unity ML Agents Training

Hi creators,

Thank you so much for creating this amazing tool for us to use. I am currently enrolled in a master of Architecture program and I have decided to use ML-Agents as my main driver for the project. With this new devotion, I am expecting to upgrade my machine.

I have read that ML-Agents do not require GPU even training could be done by only the CPU. I have some questions regarding this:

  1. Would a better GPU in ML-Agents than running it on the CPU? I am expecting myself to upgrade to NVDA RTX 2070 Super with AMD Ryzen 5 3600/
  2. Since the training option is binary: GPU or CPU, would that mean I could just invest in one CPU better than having a better GPU?
  3. What would you suggest for me to upgrade, or, maybe for the good for all, you could suggest any specific combination of hardware for a range of budget?

I am currently using a laptop with Intel 6700HQ and NVDA GTX 960 M and all my simple agent training ran pretty well so far.

Thank you so much!

AFAIK, GPUs outperform CPUs for most ML use cases. Since ML-Agents works with Tensorflow and CUDA and doesn’t have any other unusual requirements, I’d assume Tim Dettmers’ great guides would apply here too.

Given the cost for one or more high-end GPU(s), an alternative might be cloud training.

EDIT: Please see GPU is no faster for training, even with visual observations? · Issue #4129 · Unity-Technologies/ml-agents · GitHub
regarding GPU vs CPU training.

I set up ML agents with GPU (1080 gtx) with almost no effect. GPU support seems to be poor at this point.