On the contrary, I think inference is going on-device more in the future, but ‘users’ will still need decent CPUs and GPUs. Intel is well set up for this: they have good CPU, GPU, and NPU IP.
Intel can go ARM if they want, no problem, just like AMD can (and almost tried). They could theoretically preserve most of their core design and still switch ISA.
Servers will still need CPUs for a long time.
As for GPU compute, we are both in a bubble, and at several forks in the road:
-
Is bitnet ML going to take off? If it does, that shifts the advantage to almost cyptominer-like ASICs, as expensive matrix multiplication no longer matters for inference.
-
Otherwise, what about NPUs? Huawei is already using them, and even training good production models with them. Intel can try their hand at this game again if loads start shifting away from CUDA.
-
Otherwise, they still have a decent shot at the CUDA ecosystem via ZLUDA and their own frameworks. Training and research will probably forever be Nvidia (and some niches like Cerebra’s), but still.