#
✕ nick:arnd
#asahi · 2021-02-01 15:52
arndjix: do you have any idea how many AMX units there are? I have this suspicion that it's less than one per core, possibly only one of them in total, which would explain why it's slow to copy through L1 cache, and why it needs an explicit enable/disable instruction
#asahi · 2021-02-01 15:44
arndANE seems to be its own processor core with shared memory, while AMX is the instruction set extension for the CPU itself and operates on vector registers
#asahi · 2021-02-01 15:43
arndIt looks like ANE (apple neural engine) is bfloat16 and float32, while AMX is int16/float32/float64
#asahi · 2021-02-01 15:21
arndI tried to get it working inside of a linux guest running in qemu/hvf, but the instruction to enable it gives invalid-opcode in a VM
#asahi · 2021-02-01 15:19
arndsegher: https://gist.github.com/dougallj/7a75a3be1ec69ca550e7c36dc75e0d6f, TL;DR essentially a vector FMA unit to speed up 8x8 DGEMM or 16x16 SGEMM matrix operations