Correctable vs uncorrectable ECC errors

A correctable error (CE) is a memory error that ECC detected and fixed. An uncorrectable error (UE) is one above the code's correction threshold, so the system could not repair it, per the Linux RAS guide. NVIDIA's nvidia-smi guide documents single-bit errors as corrected by hardware and double-bit errors as detected but not corrected, which it calls SEC-DED. NVIDIA abbreviates single-bit errors as SBEs.

The memory controller stores CE and UE information in registers that the Linux EDAC driver reads. Intel's server board guide works from the System Event Log and treats fewer than 10 correctable events per 24 hours at one DIMM location as within threshold. NVIDIA GPUs from Ampere on use row remapping, up to 512 remaps against 64 page retirements before, applied after a GPU reset. nvidia-smi shows remapped rows plus pending and failure flags. XID 48 marks an uncorrectable ECC error.

Alert on any UE. Trend CE counts per DIMM slot, since CE events can predict later UEs, and replace a DIMM whose count keeps climbing after a reseat. On a GPU, read the remap pending and failure flags after any XID 48.

Sources

Last verified October 30, 2026.