Well, whether it matters depends on the workload: IB is basically remote DMA so if you need to pick and poke remote data I guess it'll work as another NUMA tier

Well, whether it matters depends on the workload: IB is basically remote DMA so if you need to pick and poke remote data I guess it'll work as another NUMA tier. But for AI training, where you're simply shuffling around large stacks of matrices, my guess is latency constraints weaken.