Curve data files

Curve data travels as JSON documents, so you can plot it, archive it, or move it between installs. There are two kinds:

• a reference curve — the card’s full stock V/F table, as set-reference-curve saves it;
• a tuning — the offsets a tuning run applies: the tuned curve anchors and the memory offset.

Exporting and importing

# Save the reference and also export it:
simple-nvidia-undervolt set-reference-curve --out-curve-file stock.json
# Restore that reference later (validated against the card - same GPU, same curve anchors):
simple-nvidia-undervolt set-reference-curve --in-curve-file stock.json
# Tune and also export the resulting tuning:
simple-nvidia-undervolt --mv 900 --out-tuning-file tuning.json
# Export whatever tuning is currently applied (works for a foreign one, e.g. Afterburner's):
simple-nvidia-undervolt status --out-tuning-file tuning.json
# Re-apply an exported tuning exactly, offsets as data - no planning:
simple-nvidia-undervolt --in-tuning-file tuning.json

With --dry-run, --out-tuning-file exports the planned tuning without applying anything.

Reference curve

The full stock table, one { mv, mhz } entry per anchor, plus where it came from:

{
  "type": "referenceCurve",
  "formatVersion": 1,
  "appVersion": "1.0.4",
  "savedAt": "2026-07-25 09:14",
  "gpuName": "NVIDIA GeForce RTX 5090",
  "gpuPciIds": "2B8510DE-89EE1043-000000A1-00002B85",
  "tempC": 30,
  "curve": [
    { "mv": 900, "mhz": 2392 },
    { "mv": 910, "mhz": 2415 },
    … one entry per anchor, the whole stock table …
    { "mv": 1235, "mhz": 3187 },
    { "mv": 1240, "mhz": 3195 }
  ]
}

Capture it stock, idle and cool: the reference exists to make tuning reproducible, and a curve captured tuned or hot bakes those conditions into every plan made from it. Importing checks the identity fields and the anchor voltages against the live card.

Tuning

The tuned range of anchors — offset (MHz from stock) is what a re-apply writes, mhz (the anchor’s clock at the moment of tuning) is informative — plus memoryOffset (MHz from the factory base clock):

{
  "type": "tuning",
  "formatVersion": 1,
  "appVersion": "1.0.4",
  "savedAt": "2026-07-25 09:14",
  "gpuName": "NVIDIA GeForce RTX 5090",
  "gpuPciIds": "2B8510DE-89EE1043-000000A1-00002B85",
  "memoryOffset": 1400,
  "curve": [
    { "mv": 890, "offset": 480, "mhz": 2820 },
    { "mv": 895, "offset": 480, "mhz": 2850 },
    { "mv": 900, "offset": 480, "mhz": 2872 },
    { "mv": 910, "offset": 473, "mhz": 2888 },
    … the flattened anchors above the cap …
    { "mv": 1240, "offset": -307, "mhz": 2888 }
  ]
}

Re-applying matches each entry to the card’s table by voltage and writes the offsets exactly as validated — no re-planning. Anchors the document doesn’t name stay stock. Identity is model-level on purpose: chips of the same model bin differently, but offsets land on the target card’s own stock curve, so sharing a tuning between cards of the same model is part of what the format is for — a structurally different curve table is refused by the anchor matching.

Hand-written documents

Only curve is required; every other field may be omitted. The minimal tuning — raise the 900 mV anchor by 100 MHz, leave everything else stock:

{
  "curve": [
    { "mv": 900, "offset": 100 }
  ]
}

Each tuned anchor needs mv and offset and/or mhz — given only mhz, the offset resolves as mhz - stock against the saved reference (or a live stock read). The minimal reference curve is the same document as above with the metadata dropped — but its curve must still be the complete stock table, so in practice start from an --out-curve-file export.

The optional fields degrade predictably: a document naming no GPU applies with a warning instead of an identity check, and an absent type/formatVersion is accepted (the flag a file is handed to already says which kind it holds). What must match, must match exactly: a named identity field, a present type, the anchor voltages. Unknown fields are refused like unknown CLI flags — a typo’d memoryOffest must fail, not silently change the tuning — and every offset is held to the same plausibility bounds a planned tuning gets. Anything only a foreign tool sets — a voltage boost, say — is outside the format, so it is neither exported nor re-applied.

The app’s own store files

The same documents are how the app stores its own state, under data\ in the install directory (%ProgramFiles%\simple-nvidia-undervolt):

• data\reference-curve.json — the saved reference curve (the same document --out-curve-file exports).
• data\persisted-tuning.json — the tuning the logon task re-applies (the same document --out-tuning-file exports). It lives under Program Files deliberately: the file is admin-only writable, and the elevated logon task must not consume user-writable data. clear removes it along with the task.