alt Hacker NewsWider SIMD would be useful, especially with AVX-512 style improvements. 1024 or even 2048 bits wide operations.
Of course memory bandwidth should increase proportionally otherwise the cores might have no data to process.
Replies
owlbite • yesterday at 7:54 PM
Much better to burn the area for multiple smaller units, its a bit more area for frontend handling, but worth it for the flexibility (see Apple's M-series chips vs intel avx*).
➕ show 1 reply
TimorousBestie • yesterday at 7:26 PM
I would love to be able to fit small matrices (4x4 or 16x16 depending on precision) in SIMD registers together with intrinsics for matrix arithmetic.
account4mypc • yesterday at 11:41 PM
AMX registers are 1024 *bytes*
I wouldn't mind, but might need to increase cache line size on x86, as avx512 has reached the current size.