Comment by wuubuu

Randomized matrix sketching is one way to get at this (see https://arxiv.org/abs/2302.11474), the problem is hardware is heavily optimized for dense multiplies so what you save in flops doesn't translate to real runtime speeds ups.