Hi! I learnt that in A100 exist a “MIG” technique to divide one GPU into separate small GPUs, and I am wondering can they communicate with each other? Thanks!!
For MIG on A100 and A30, can different GPU instance communicate?
Ordinary communications like host-based MPI are possible. Also, a limited (between CI but not between GI) form of CUDA IPC is possible. P2P is not possible between MIG partitions, currently.