I think the distinction with ML inference workloads is that you often have very little control flow, so this type of architecture lets you match layers to adjacent cores so that each operation gets it's data directly from the step before rather than from RAM.