logoalt Hacker News

WithinReason • today at 11:00 AM • 1 reply • view on HN

Wouldn't a SIMD/SIMT architecture achieve similar speedups? There are many ways to parallelise programs. What am I missing?


Replies

Tuna-Fish • today at 11:26 AM

You need either the compiler or the programmer to do the parallelization, and in practice the communication costs between cores can be large enough for the parallelization to not be worth it for simple programs.

This ISA seems to be designed to: 1. make the automatic parallelization easy (with workload split between compiler and hardware) 2. make communication costs between tiles very small, so that even very simple workloads can be automatically parallelized.

➕ show 1 reply