Zubnet AILearnWiki › Flash Attention
Infrastructure

Flash Attention

FlashAttention, FlashAttention-2
A GPU-optimized implementation of the attention mechanism that is 2–4x faster and uses significantly less memory than standard attention. Flash Attention achieves this not by changing what attention computes, but by restructuring how the computation is performed on GPU hardware — minimizing slow memory transfers between GPU HBM and on-chip SRAM.

Why it matters

Flash Attention is arguably the most impactful systems optimization in modern AI. It made long-context models practical by reducing attention's memory usage from quadratic to near-linear (in practice), directly enabling the jump from 4K to 128K+ context windows. Every major LLM uses it. Without Flash Attention, today's long-context models would be prohibitively expensive.

Deep Dive

The key insight (Dao et al., 2022): standard attention materializes the full N×N attention matrix in GPU HBM (high bandwidth memory), which is both memory-intensive (quadratic in sequence length) and slow (HBM bandwidth is the bottleneck). Flash Attention never materializes this matrix. Instead, it computes attention in tiles, loading small blocks of Q, K, V into fast on-chip SRAM, computing partial results, and accumulating them — a technique called "tiling" or "kernel fusion."

The Memory Savings

Standard attention stores the N×N attention matrix, requiring O(N²) memory. For a 128K context with 128 attention heads, that's hundreds of gigabytes. Flash Attention uses O(N) memory by computing softmax incrementally and never storing the full matrix. This is what made 128K–1M context windows feasible on existing hardware. FlashAttention-2 further improved throughput by better parallelizing across GPU thread blocks.

IO-Aware Algorithm Design

Flash Attention exemplifies a broader principle: on modern hardware, the bottleneck is often memory bandwidth, not compute. GPUs can perform trillions of operations per second but can only read/write hundreds of gigabytes per second from HBM. Algorithms that minimize memory traffic (even at the cost of extra computation) often win. This "IO-aware" approach is influencing how the entire field thinks about algorithm design for AI workloads.

Related Concepts

← All Terms
← Fine-tuning FLOPs →