{
  "schema_version": "1.2.0",
  "rubric_version": "1.1",
  "license": "CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)",
  "record": {
    "id": "O-0121",
    "slug": "pearl-research",
    "title": "Pearl Research Labs",
    "aliases": [],
    "status": "published",
    "last_reviewed": "2026-09-25",
    "review_interval_days": 90,
    "steward": null,
    "provenance": {
      "drafted_by": "ai",
      "reviewed_by": [
        "codex-review"
      ]
    },
    "risk_flags": [],
    "flags": [
      "provider-reported"
    ],
    "kind": "company",
    "homepage": "https://pearlresearch.ai/",
    "one_liner": "A company building AI inference infrastructure and Pearl, a blockchain whose proof-of-useful-work mining is a by-product of GPU matrix multiplication.",
    "sources": [
      {
        "source": "S-3603",
        "supports": "infrastructure and algorithms for AI inference"
      },
      {
        "source": "S-1105",
        "supports": "FP8 proof-of-useful-work protocol specification"
      },
      {
        "source": "S-1106",
        "supports": "integer matrix-multiplication whitepaper; mining as a side-effect of AI workloads"
      },
      {
        "source": "S-1107",
        "supports": "public monorepo: full node, GPU miner, proof-of-work circuit and verifier; README cites the Komargodski-Weinstein proposal"
      },
      {
        "source": "S-1609",
        "supports": "proof of useful work from arbitrary matrix multiplication (Komargodski and Weinstein)"
      },
      {
        "source": "S-1010",
        "supports": "Hawkeye (MLSys 2026): two of four authors at Pearl Research Labs; bit-exact CPU replay of GPU matrix multiplication"
      }
    ],
    "type": "organization",
    "url": "https://trustbutveri.fyi/organizations/pearl-research/",
    "source_file": "content/organizations/pearl-research.md",
    "flags_all": [
      "provider-reported"
    ],
    "body_markdown": "Pearl Research Labs describes itself as building infrastructure and algorithms for AI inference [[S-3603]]. It develops the [[I-0004|Pearl blockchain]], an instance of [[M-0007|proofs of useful work]]:\n\n- **Protocol specification.** Pearl specifies a proof-of-useful-work protocol in which the unit of mining work is FP8 matrix multiplication on GPUs [[S-1105]].\n- **Integer whitepaper.** Its earlier whitepaper presents mining as a side-effect of the matrix multiplications in AI training and inference workloads [[S-1106]].\n- **Open-source network.** Pearl publishes a monorepo with a reference full node, a vLLM-based GPU miner and a proof-of-work circuit and verifier [[S-1107]]. Its README says the mining follows the proof of useful work from arbitrary matrix multiplication proposed by Komargodski and Weinstein [[S-1107]] [[S-1609]].\n- **Reproducible GPU arithmetic.** Two Pearl researchers, with two Stanford co-authors, published Hawkeye at MLSys 2026. It re-executes on a CPU, bit for bit, the matrix multiplications that NVIDIA Ampere, Hopper and Ada Lovelace GPUs perform in FP16, BF16 and FP8 [[S-1010]]; see [[M-0002]].",
    "body_text": "Pearl Research Labs describes itself as building infrastructure and algorithms for AI inference [S-3603]. It develops the Pearl blockchain, an instance of proofs of useful work: - Protocol specification. Pearl specifies a proof-of-useful-work protocol in which the unit of mining work is FP8 matrix multiplication on GPUs [S-1105]. - Integer whitepaper. Its earlier whitepaper presents mining as a side-effect of the matrix multiplications in AI training and inference workloads [S-1106]. - Open-source network. Pearl publishes a monorepo with a reference full node, a vLLM-based GPU miner and a proof-of-work circuit and verifier [S-1107]. Its README says the mining follows the proof of useful work from arbitrary matrix multiplication proposed by Komargodski and Weinstein [S-1107] [S-1609]. - Reproducible GPU arithmetic. Two Pearl researchers, with two Stanford co-authors, published Hawkeye at MLSys 2026. It re-executes on a CPU, bit for bit, the matrix multiplications that NVIDIA Ampere, Hopper and Ada Lovelace GPUs perform in FP16, BF16 and FP8 [S-1010]; see Deterministic and bit-exact inference.",
    "referenced_by": [
      {
        "id": "M-0007",
        "title": "Proofs of useful work for capacity accounting",
        "url": "https://trustbutveri.fyi/mechanisms/proofs-of-useful-work/"
      },
      {
        "id": "I-0004",
        "title": "Pearl proof-of-useful-work blockchain",
        "url": "https://trustbutveri.fyi/implementations/pearl-proof-of-useful-work/"
      },
      {
        "id": "S-1010",
        "title": "Hawkeye: Reproducing GPU-Level Non-Determinism",
        "url": "https://trustbutveri.fyi/sources/badash-hawkeye/"
      },
      {
        "id": "S-1105",
        "title": "Pearl Floating Point Scheme Specification",
        "url": "https://trustbutveri.fyi/sources/pearl-floating-point-scheme-specification/"
      },
      {
        "id": "S-1106",
        "title": "Pearl INT Whitepaper",
        "url": "https://trustbutveri.fyi/sources/pearl-int-whitepaper/"
      },
      {
        "id": "S-1107",
        "title": "pearl: Monorepo for the Pearl network",
        "url": "https://trustbutveri.fyi/sources/pearl-network-monorepo/"
      },
      {
        "id": "S-3571",
        "title": "Proof of Useful Work from the Ground Up",
        "url": "https://trustbutveri.fyi/sources/pearl-proof-of-useful-work-blog/"
      },
      {
        "id": "S-3603",
        "title": "Pearl Research Labs homepage",
        "url": "https://trustbutveri.fyi/sources/pearl-research-homepage/"
      }
    ]
  }
}