{
  "abstract": "While most attention logits can be computed in low precision without degrading numerical stability, current attention kernels fail to exploit this phenomenon. We introduce a novel hardware-algorithm co-design in the form of mixed-precision FlashAttention. Our method accumulates key-query products and evaluates their exponentials in 8-bit formats, then adaptively identifies sensitive sub-blocks and recomputes them in 16-bit formats. We propose the specifications for a dedicated accelerator capable of executing this pipeline efficiently. Simulated experiments with Qwen3 and Gemma 3 show that rerouting a selective minority of sub-blocks to high precision is sufficient to recover the baseline model performance.",
  "abstract_shown": true,
  "affiliations": [],
  "article_type": null,
  "arxiv": "2609.39361",
  "author_total": 8,
  "authors": [
    {
      "name": "Stanislav Budzinskiy"
    },
    {
      "name": "Marian Gloser"
    },
    {
      "name": "Tolunay Yilmaz"
    },
    {
      "name": "Ying Hong Tham"
    },
    {
      "name": "Yuanyi Lin"
    },
    {
      "name": "Wenyi Fang"
    },
    {
      "name": "Fan Wu"
    },
    {
      "name": "Philipp Petersen"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/7x56kyb8bt1yr.bib",
  "cited_by": 0,
  "doi": "10.48550/arxiv.2609.39361",
  "fulltext_url": null,
  "id": "7x56kyb8bt1yr",
  "issns": [],
  "landing_url": "https://arxiv.org/abs/2609.39361",
  "language": null,
  "licence": "CC-BY-4.0",
  "licence_label": "CC BY 4.0",
  "licence_source": "article",
  "licence_url": "http://creativecommons.org/licenses/by/4.0/",
  "metadata_licence": "CC0-1.0",
  "native_subjects": [
    [
      "arxiv",
      "cs.LG"
    ],
    [
      "arxiv",
      "cs.NA"
    ],
    [
      "arxiv",
      "math.NA"
    ],
    [
      "openalex",
      "1708|Parallel Computing and Optimization Techniques"
    ]
  ],
  "openalex": "W7215098400",
  "pdf_url": "https://arxiv.org/pdf/2609.39361",
  "pmcid": null,
  "pmid": null,
  "published": "2026-09-30",
  "publisher": null,
  "ris": "https://research.draconitic.com/paper/7x56kyb8bt1yr.ris",
  "sources": [
    {
      "first_seen": "2026-10-30T12:25:49Z",
      "native_id": "2609.39361",
      "source": "arxiv",
      "url": "https://arxiv.org/abs/2609.39361"
    }
  ],
  "subjects": [
    {
      "field": "mathematics",
      "origin": "arxiv:math.NA",
      "subfield": "numerical-analysis"
    }
  ],
  "title": "LampAttention: Look-Ahead Mixed-Precision FlashAttention for Dedicated Accelerators",
  "title_text": "LampAttention: Look-Ahead Mixed-Precision FlashAttention for Dedicated Accelerators",
  "type": "preprint",
  "updated_at": "2026-10-30T12:25:49Z",
  "url": "https://research.draconitic.com/paper/7x56kyb8bt1yr",
  "venue": "arXiv",
  "versions": []
}