Overview
SIMD is a form of data-level parallelism. It is particularly effective for workloads that involve performing the same operation on large arrays of data, such as image processing, video encoding, and scientific simulations.
Examples
- Intel/AMD: MMX, SSE, AVX, AVX-512.
- ARM: NEON, SVE.
Mechanism
Instead of adding two numbers, a SIMD instruction might add two vectors of four or eight numbers each in a single clock cycle using wide registers.