Yes, there are some fairly revolutionary(-ish) chips. Those are few and far between because they tend to be hyper specialized. Inference but not training or only optimized for a very small input matrix (common for edge computing like cameras).
By and large? They really ARE “traditional” GPGPUs that are optimized to hell and back for vector operations and linear algebra. And a lot of the gains there come from multiplying their floating point performance by 2-4 (depending on if half or quarter precision). They aren’t as good for double precision as something optimized for it but basically only a very small subset of users need that. There will be no issues repurposing the hardware in these data centers.
And the rest is data movement which has always been the real problem.
Ehhhh.
Yes, there are some fairly revolutionary(-ish) chips. Those are few and far between because they tend to be hyper specialized. Inference but not training or only optimized for a very small input matrix (common for edge computing like cameras).
By and large? They really ARE “traditional” GPGPUs that are optimized to hell and back for vector operations and linear algebra. And a lot of the gains there come from multiplying their floating point performance by 2-4 (depending on if half or quarter precision). They aren’t as good for double precision as something optimized for it but basically only a very small subset of users need that. There will be no issues repurposing the hardware in these data centers.
And the rest is data movement which has always been the real problem.