The same discrepancies that apply to software benchmarking programs also apply to predictions of hardware failure rates based on subjecting them to extreme stress conditions.  Much like a graphics card, the only way to know how it will perform is to actually use it in real life conditions for hours, days or years.  Such is the case with DRAM errors, as proven by a study done by Google.  Instead of the current estimate based on synthetic testing, as Ars Technica puts it, ‘DRAM errors are not 200 to 5,000 failures in time per billion hours of operation (FIT) per Mbit; Google found that their numbers were between 25,000 and 75,000 FIT per Mb‘.   The same caveats apply as they do for the mean time before failure rating of a hard drive, but you could expect to see just over 3 errors per minute on a system with 4GB of DRAM, most likely stemming from a single chip around its 2nd birthday.

“The conventional wisdom about DRAM error rates is that errors are rare, and the majority of the errors that do occur are so-called “soft errors” – randomly corrupted bits that have been flipped by incoming cosmic rays. But a recent large- scale study of DRAM errors released by Google turns this wisdom on its head, and in doing so reinforces the importance of error correction coding (ECC) and regular hardware replacement for datacenter machines. ”

Here is some more Tech News from around the web:

Tech Talk