← Back to reference

Supported Dcgm Matrix

DCGMTestedNotes
3.2YesSame, minus 8 fields
3.1Best effortNo PCIe link width bug
2.4NoOld API
4.0 (preview)Best effortNew field IDs

Required APIs

  • dcgmInit / dcgmShutdown
  • dcgmStartEmbedded (hostengine v2)
  • dcgmGroupCreate + dcgmGroupAddEntity
  • dcgmGetLatestValuesForFields

Field IDs Implemented

See agents/denpex-enterprise-agent/src/gpu/dcgm_fields.rs.

Every DCGM_FI_DEV_* listed in Phase 6.2 of the spec is mapped to GpuMetric.

Diagnostic Levels

  • Level 1 (Short) — ~30 s run, ~5 fields.
  • Level 2 (Medium) — ~2 min, ~20 fields.
  • Level 3 (Long) — ~10 min, full sweep.

All three call dcgmRunDiagnostic and emit the results.

MIG Slices

Per-instance via dcgmGroupAddMigInstance.

Compatibility Notices

  • DCGM 3.0 (and older) does not expose PCIe replays correctly on A30 — recommend upgrading.
  • DCGM 3.3 sometimes returns stale values during heavy load; document under docs/operations/dcgm-staleness.md.