Deep Learning Custom Rig with Titan V and 1080 Ti

I have a doubt regarding building a custom rig. Our institute has money for 2 titan v s and 6 1080Tis. So is it possible to create a rig with this combination of different kind of GPUs?
Has anyone done that before?

It can work but it depends for what. I personally have tried with my GTX 1080Ti and GTX 1070 and here are the possible use case where it can work/doesn’t work:

  • Cross validation: can work. If your batch size doesn’t exceed the maximum amount of Vram of your card with the least memory then you’ll get faster results. Ex: a 6 fold CV with each fold running on a GTX 1080Ti will be faster (imo) than a 6 fold CV on 2 titan V.
  • Distributed mini batch: won’t work well. As your results will have to be synchronized across all your cards the overall training will be slower.

Apart from these example you have to tell you these two things:

  • Getting 6 GTX 1080Ti working on the same motherboard will inevitably force you to buy an entreprise grade motherboard (I know no consumer grade mb with 6x 16x PCI-E slots)
  • Getting 6 GTX 1080Ti will consume much more power and you may end up with a lot more hardware failures
  • The GTX Titan V uses the volta architecture which has been tailored for deep learning with tensor cores, more fp16 compute capabilities and faster memory (GTX 1080Ti uses Pascal architecture)
  • etc…

So if I had the same choice as you I would opt for the 2 titan V without a doubt.