> Do you really need a fully connected mesh? Doesn't Thunderbolt just show up as a network connection that RDMA is ran on top of?
If you daisy chain four nodes, then traffic between nodes #1 and #4 eat up all of nodes #2 and #3's bandwidth, and you eat a big latency penalty. So, absent a switch, the fully connected mesh is the only way to have fast access to all the memory.
Obviously don't daisy chain, that wastes ports so badly. But if you connect 4 nodes into a loop, it goes fine. Relaying only adds 33% extra traffic. And what specifically are the latency numbers you have in mind?
If you have 3 links per box, then you can set up 8 nodes with a max distance of 2 hops and an average distance of 1.57 hops. That's not too bad. It's pretty close to having 2 links each to a big switch.
Sure, one could think of some kind of pipeline parallelism where you only need a fast transfer to the next step in the model and that would boost throughput but not increase model size.
If you daisy chain four nodes, then traffic between nodes #1 and #4 eat up all of nodes #2 and #3's bandwidth, and you eat a big latency penalty. So, absent a switch, the fully connected mesh is the only way to have fast access to all the memory.