ECC memory
Memory that can detect and correct certain data errors on its own. Used in servers, workstations and data center GPUs.
ECC stands for Error-Correcting Code. In addition to the data, ECC memory stores a few check bits that make it possible to verify every read. If a bit has been flipped, the memory corrects it on the fly; if two bits are affected, it reports the error instead of passing on a wrong value.
These errors are rare, but they do happen: cosmic radiation, heat, an ageing component. On a desktop PC they go unnoticed or cause the occasional crash. On a machine that runs around the clock or trains an AI model for days, a single silent error can corrupt a result or waste hours of computation. That is why servers, workstations and data center GPUs, with their HBM memory, use ECC.
For system memory, ECC requires a compatible processor and motherboard, such as Threadripper, EPYC or Xeon platforms, and suitable modules, often RDIMMs. Note that consumer DDR5 includes on-die error correction inside each chip, which does not replace true end-to-end ECC. On the GPU side, professional cards let you enable ECC on their memory, sometimes at the cost of a slight reduction in capacity or speed. On a product page, look for the ECC mention and the module type.
Our Threadripper AI workstation is fitted with DDR5 ECC memory, designed for long computing sessions.