{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-04-014",
    "source_story_id": "tmp-story-fp4-flashattention-four",
    "edition_id": "mp-2026-09-04-morning-0057",
    "edition_url": "https://themachinepress.com/edition/2026-09-04",
    "position": 14,
    "story_type": "dispatch",
    "section": "chips-infrastructure",
    "editorial_classification": "editorial",
    "headline": "Four-Bit Matrix Math Was Waiting on the Softmax",
    "slug": "four-bit-matrix-math-was-waiting-on-the-softmax",
    "dek": "A hardware-aware FlashAttention path maps scores directly to FP4 probabilities and carries forward quantization into the backward pass.",
    "summary": "A hardware-aware FlashAttention path maps scores directly to FP4 probabilities and carries forward quantization into the backward pass.",
    "body_text": "Shrinking matrix products on Blackwell tensor cores exposes conversion and on-chip dependency costs elsewhere in attention. Direct-P reached up to 2.13 times BF16 forward throughput for noncausal inference on an NVIDIA GB200. A causal training route reused quantized queries and keys with FP8 gradient operands, accelerating a complete single-GPU 8B update by up to 1.14 times. In matched distributed training, every tested MXFP4 probability-and-value trajectory diverged, sharply limiting the claim.",
    "why_it_matters": "A hardware-aware FlashAttention path maps scores directly to FP4 probabilities and carries forward quantization into the backward pass.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-04-014/four-bit-matrix-math-was-waiting-on-the-softmax",
    "json_url": "https://themachinepress.com/story/mp-2026-09-04-014.json",
    "first_published_at": "2026-09-04T09:00:00.000-04:00",
    "modified_at": "2026-09-04T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-04-014-001",
        "text": "A hardware-aware FlashAttention path maps scores directly to FP4 probabilities and carries forward quantization into the backward pass.",
        "source_ids": [
          "source-2026-09-04-014"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-04-014"
    ],
    "tags": [
      "FP4",
      "FlashAttention",
      "Blackwell"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-04-014",
      "title": "arXiv preprint 2609.04105",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.04105",
      "canonical_url": "https://arxiv.org/abs/2609.04105",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-03T13:12:35.000-04:00",
      "accessed_at": "2026-09-04T08:24:03.294-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-04-014-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Four-Bit Matrix Math Was Waiting on the Softmax",
    "publisher": "The Machine Press",
    "published_at": "2026-09-04T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-04-014/four-bit-matrix-math-was-waiting-on-the-softmax"
  }
}