Linus Torvalds on Why ARM Won't Win the Server Space


The fragmentation of the arm space is a concern to me, especially the sort of hijinks the likes of Apple are reportedly doing with their chips to speed up certain ops; but god knows what they're really doing because they don't publish their ABIs. But in practice, if something compiles under gcc or clang then porting between x86 and some real world available ARM platform is usually rather straightforward if you're not using 3d graphics.

I can see the dependency on nvidia's cuda architecture for lots of deep learning research being problematic though. You can simulate a slower version of a fairly complex deep network using a modern video card plugged into a PCIe bus, but trying hooking any of that up to an ARM cpu.
 
Back
Top