Exactly. First there was nvswitch, which dramatically increased the bandwidth over pcie. But that didn't scale to a large number of GPUs. Then there was nvswitc

Exactly. First there was nvswitch, which dramatically increased the bandwidth over pcie. But that didn't scale to a large number of GPUs. Then there was nvswitch, which solved the scaling problem inside a node. I wouldn't be surprised if the next leap is something like nvlink cables between nodes that don't need traditional routing capabilities.