FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision
Summary
FlashAttention-3 speeds up attention on Hopper GPUs with three main techniques: overlapping computation and data movement via warp-specialization, interleaving block-wise matmul and softmax, and block quantization plus incoherent processing that uses hardware support for FP8. They report only 35% utilization on the H100 GPU. They report that FlashAttention-3 achieves a 1.5-2.0x speedup on H100 GPUs, with FP16 reaching up to 740 TFLOPs/s (75% utilization) and FP8 reaching close to 1.2 PFLOPs/s. They also report that FP8 FlashAttention-3 achieves 2.6x lower numerical error than a baseline FP8 attention. [S1]
Claims
- id: C1 claim: 'The authors say FlashAttention-2 achieved only 35% utilization on the H100 GPU.' source: S1 evidence: "FlashAttention-2 achieving only 35% utilization on the H100 GPU"
- id: C2 claim: 'They report a 1.5-2.0x speedup on H100, with FP16 reaching up to 740 TFLOPs/s (75% utilization).' source: S1 evidence: "with FP16 reaching up to 740 TFLOPs/s (75% utilization)"
- id: C3 claim: 'They report that FP8 FlashAttention-3 reaches close to 1.2 PFLOPs/s.' source: S1 evidence: "with FP8 reaching close to 1.2 PFLOPs/s"
Verification
- Code available: Paper: https://github.com/Dao-AILab/flash-attention, which returned HTTP 200 on 2026-11-12. Not run.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: Live notes 2205-14135-flashattention and 2307-08691-flashattention-2. This is the paper-verification note for FlashAttention-3 itself.
- Limits: Speed and utilization figures are author-reported on H100 GPUs (v2, 12 Jul 2024). This note does not re-benchmark the kernel. The abstract describes a 1.5-2.0x speedup range for the H100 results.
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 11 Jul 2024 (v1); last revised 12 Jul 2024 (v2). Sources: [S1] https://arxiv.org/abs/2407.08608, v2 (12 Jul 2024). checked_on: 2026-11-12.
When to apply
Use when asking how FlashAttention was adapted to Hopper asynchrony and FP8, or when comparing FA-2 versus FA-3 utilization claims.
Facts
- id: F1 claim: 'FlashAttention-3 targets Hopper GPUs with asynchrony, warp-specialization, and FP8 low-precision techniques.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F2 claim: 'The authors report that FlashAttention-2 reached only 35% utilization on H100, versus up to 75% for FP16 FlashAttention-3.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F3 claim: 'They report up to 740 TFLOPs/s in FP16 and close to 1.2 PFLOPs/s in FP8, with 2.6x lower numerical error than a baseline FP8 attention.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
Sources
- [S1] arXiv 2407.08608, https://arxiv.org/abs/2407.08608. License: CC BY 4.0. v2 (12 Jul 2024). Accessed 2026-11-12.
License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.
License: CC-BY-4.0 · Summary by iinventedlight (AI-assisted). Source: Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao, FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision (CC BY 4.0).
Cite this
'iinventedlight (AI-assisted)'. 2026. "FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision." iinventedlight Library, iil:note:vnk0qjet, version of 2026-11-12. https://iinventedlight.com/library/notes/2407-08608-flashattention-3/. License: CC-BY-4.0.
@misc{iil_note_vnk0qjet,
author = {{'iinventedlight (AI-assisted)'}},
title = {{FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision}},
howpublished = {iinventedlight Library},
year = {2026},
date = {2026-11-12},
url = {https://iinventedlight.com/library/notes/2407-08608-flashattention-3/},
note = {Permanent ID iil:note:vnk0qjet. License: CC-BY-4.0.},
}
Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.
For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2407-08608-flashattention-3.md.