it's Kanbaru again 🌟

has the size of the register file on a modern cpu surpassed the size of the n64 gpu's texture cache (4KiB) yet. it must be at least halfway right

FCLC

(replying to it's Kanbaru again 🌟)

@hikari yes, many times over!

The per thread addressable register file per core is:
AVX512(f) vector RF 32x512b -> 2 KiB
AVX512(BW) mask RF 8x64b -> .064 KiB
AMX matrix RF 8x8096b -> 8 KiB
APX GPR 32x64b -> .256 KiB

That ignores that current physical register files tend to be between 2-3X the software addressable RF; helps for performance when context switching by not having to pop all of register state to cache

FCLC

(replying to FCLC)

@hikari So roughly 10.3 KiB addressable, ~25 KiB physical.

Assume relatively small server socket at 32C, that’s ~0.8MiB of register file per socket.

Saagar Jha

(replying to FCLC)