Skip to content
Discussion options

You must be logged in to vote

Yes. Three levels:

  1. AMD works: the DirectML backend path runs AMD GPUs/iGPUs directly.
  2. Mixed-vendor multi-GPU: CUDA groups and DML groups coexist in one process; chains are pinned to groups (chain % n_groups), so every game always hits the same backend/kernel set — reruns stay bitwise identical across a mixed fleet.
  3. Real gains depend on model size: on a light MLP workload, dGPU + iGPU measured 1.8× a single card (1113–1138 vs 602–683 games/s). On heavy models a slow iGPU contributes little — profile first.

Also note: you don't need a second card at all. Splitting one GPU into multiple device groups is official usage (1 → 6 groups: 780 → 2303 games/s measured).

Replies: 1 comment

Comment options

You must be logged in to vote
0 replies
Answer selected by SkYContact
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Category
Q&A
Labels
None yet
1 participant