Ok, what FlashAttention changes is space complexity: from O(N^2) to O(N). Time complexity is still ~O(N^2) as with standard Self-Attention.

In other words, optimizes practical runtime through I/O reduction without altering asymptotic complexity