ECC and DDR5

The strongest technical point people added is that the distinction is not philosophical. On-die ECC and system ECC cover different failure modes. Several comments stressed that DDR5 on-die ECC mainly masks internal DRAM cell errors and leaves transmission issues, socket problems, board issues, and some mechanical faults untouched. That turned the conversation away from “DDR5 already has ECC” and toward “what errors can you actually see and act on.” A recurring example was memory that works for months, then starts throwing correctable errors in one module. With real ECC and BMC or EDAC logs, you can identify the failing stick or a bad slot. Without that visibility, the same machine may run for years while occasionally writing wrong data. People were far less aligned on frequency and severity. Some reported corrected errors every few months or after moving a machine. Others said they have run ECC-equipped systems for years and never seen a single correction. The usable consensus was narrower than the rhetoric. Healthy systems at rated settings should not log frequent corrections. If they do, suspect a bad DIMM, poor seating, bent socket pins, power issues, overclocking margins, or aging hardware rather than treat it as normal background noise. Cosmic radiation came up often, but several comments pushed back on the lazy version of that story. Bit flips do happen from radiation, especially as densities rise and altitude increases, but many real-world errors are correlated hardware faults, contact problems, or timing instability. A second thread focused on the lack of current public data. People wanted fresh large-scale studies for DDR5, not just repeated references to the old Google DRAM paper. The practical answer was that only hyperscalers can gather that data at scale, and they likely do not publish it. That leaves everyone else with anecdotes, which are unsatisfying but still operationally useful when the anecdotes come with error logs and failed parts. The market discussion was blunt. Many comments blamed decades of CPU and motherboard segmentation for making ECC harder to buy than it should be. AMD was cited as better than Intel on raw CPU support, but that did not solve the real buyer problem: finding a compatible motherboard and affordable modules. DDR5 ECC UDIMMs were described as scarce and expensive, while used ECC RDIMMs can be cheap if you are willing to build around workstation or server hardware. That made one concrete recommendation stand out above the broader “ECC everywhere” rhetoric. If integrity matters and budget matters, second-hand enterprise platforms may be the easiest path to real ECC today. The most contentious part was the call for government action. Plenty of people liked the goal of making ECC more common, but many rejected taxes or mandates as overreach. Even supporters of wider ECC mostly landed on a simpler claim: the market is distorted, consumer buyers are poorly informed, and DDR5 on-die ECC risks making people think they already have protection they do not actually have.