There’s more to performance than peak theoretical performance. The architectures that stick around have tended to be ones with more software support (e.g., x86)

There’s more to performance than peak theoretical performance. The architectures that stick around have tended to be ones with more software support (e.g., x86). If you mean network bandwidth, the level of batching controls the bandwidth bound and is configurable. If you mean chip bandwidth, that relies on advanced compilation that is pretty darn hard to get right. For a chip like cerebras’s to win, it’d have to have a bandwidth bound workload and deliver on the software to eliminate the bottlenecks.