[fourthark]

Seems like training would be a better match, where you need tons of compute but don’t care about latency.

[ronsor]

No, the problem is that with training, you do care about latency, and you need a crap-ton of bandwidth too! Think of the all_gather; think of the gradients! Inference is actually easier to distribute.

[meehai]

Yeah, but if you can do topologies based on latencies you may get some decent tradeoffs. For example with N=1M nodes each doing batch updates in a tree manner, i.e the all reduce is actually layered by latency between nodes.