{
  "slug": "flash-cosmos",
  "term": "Flash-Cosmos",
  "aliases": [
    "Flash Computation with One-Shot Multi-Operand Sensing"
  ],
  "definition": "Flash-Cosmos is a research technique that performs bulk bitwise logic inside NAND flash using multi-wordline sensing and enhanced SLC-mode programming.",
  "note": "Mechanism: NAND strings act as AND paths when selected wordlines are sensed together; parallel strings enable OR. Enhanced SLC-mode programming widens the separation of the two voltage states with finer programming steps. This moves computation into the array rather than merely into an SSD-controller accelerator. [cosmos-talk]\n\nEvidence: The 2022 study tested 160 48-layer TLC chips from five wafers, sampling 120 blocks per chip. Its ESP test observed zero errors over more than 4.83e11 stored bits after 10,000 P/E cycles and accelerated retention representing one year at 30 C. Permanently defective cells were excluded. Footnote 11 explicitly says this does not establish the JEDEC UBER requirement. [cosmos-paper]\n\nEvaluation: Application speedups combine measured host work with Ramulator and modified MQSim models, not an end-to-end deployed SSD benchmark. Operand placement, specialized commands and firmware support are prerequisites; an NVMe label alone does not establish them. [cosmos-paper]\n\nECC boundary: XOR preserves a fixed binary linear encoding, but two individually correctable input faults can exceed the output decoder's correction radius. Correcting operands and decoding a computed result are different pipelines; a specified Hamming-code replay illustrates the distinction. [ecc-order]",
  "status": "research",
  "advantages": [
    {
      "text": "Reducing operands inside the array can avoid transferring every input over the NAND-to-controller link.",
      "citationIds": [
        "cosmos-talk"
      ]
    }
  ],
  "limitations": [
    {
      "text": "SLC-mode operands trade capacity for margin, and more precise programming takes extra time.",
      "citationIds": [
        "cosmos-talk"
      ]
    },
    {
      "text": "Operands on different chips require movement or another compute path; ordinary encryption is not transparently compatible.",
      "citationIds": [
        "cosmos-paper"
      ]
    }
  ],
  "related": [
    "computational-storage",
    "nand-flash",
    "slc",
    "mqsim",
    "parabit"
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "citations": [
    {
      "id": "cosmos-paper",
      "title": "Park et al. Flash-Cosmos. MICRO 2022, pp. 937-955, sections 5-8 and 10, footnotes 9/11. DOI: 10.1109/MICRO56248.2022.00069.",
      "url": "https://arxiv.org/pdf/2209.05566v1",
      "accessed": "2026-11-18"
    },
    {
      "id": "cosmos-talk",
      "title": "Park et al. Flash-Cosmos: MICRO 2022 author presentation, mechanisms and evaluation.",
      "url": "https://people.inf.ethz.ch/omutlu/pub/FlashCosmos_micro22-talk.pdf",
      "accessed": "2026-11-18"
    },
    {
      "id": "hesela-nvme",
      "title": "Hesela: NVM Express, computational-program support and physical implementation.",
      "url": "https://hesela.com/tech/nvme/",
      "accessed": "2026-11-18"
    },
    {
      "id": "ecc-order",
      "title": "Hesela: ECC operation order, exact Hamming-code construction, method and results.",
      "url": "https://hesela.dev/research/ecc-operation-order/README.md",
      "accessed": "2026-11-18"
    },
    {
      "id": "ecc-order-replay",
      "title": "Hesela: offline ECC operation-order replay, Python code, tests and exact results (ZIP).",
      "url": "https://hesela.dev/research/ecc-operation-order.zip",
      "accessed": "2026-11-18"
    }
  ],
  "updated": "2026-11-18",
  "provenance": [
    "Source review of the original paper and author slides, not a hardware replication. The slides describe the same work, not independent corroboration.",
    "No field error rate, universal error-free guarantee or present-day product availability is inferred.",
    "The ECC operation-order example enumerates synthetic finite populations with ideal logic and fixed layouts, not physical NAND failures. [ecc-order]"
  ]
}
