FlashAttention shows why FLOPs alone are a bad lens on GPU performance

techNmak · x · 2026-09-10

Using FlashAttention as a case study, the author explains why FLOPs are a poor proxy for GPU performance. Standard dense attention (QK^T, softmax, times V) is already highly parallel — the real pain is what naive implementations do around the math: materializing huge N×N intermediates and round-tripping them through high-bandwidth memory.

FlashAttention's core contribution is optimizing exactly those memory accesses: tiling plus online softmax keeps intermediates on-chip in SRAM instead of hitting HBM, dramatically improving real-world throughput. A great illustration that attention is memory-bound, not compute-bound.

Related event: Understanding FlashAttention: Why Memory Traffic Beats FLOPs(2 posts)→

Original post →

More from Infra

Infra channel →