> I suspect we're going to end up with a wildly different architecture before it's all over.
This is my naïveté showing but aren't GPUs (relatively) general purpose? Do "AI" workflows require that flexibility or would they be better served by special purpose hardware? To use the obvious analogy, what happened with cryptocurrency mining moving from GPU to ASIC.
Yes - I believe the shift is coming. For inference especially, check out Groq’s chip design and its simplicity and speed compared to general purpose GPUs.
This is my naïveté showing but aren't GPUs (relatively) general purpose? Do "AI" workflows require that flexibility or would they be better served by special purpose hardware? To use the obvious analogy, what happened with cryptocurrency mining moving from GPU to ASIC.