8BIT HIPPO ATTENTION MAY BE 3X FASTER THAN FLASH ATTENTION V2

Summary from the TLDR AI newsletter

A quantization-aware attention alternative that speeds up inference substantially over some state-of-the-art implementations. Its creators also claim 700 TFLOPs on MI-300X. The details are sparse, but it would be quite an achievement if so.