3 ms·
There have been many studies that look at error rates on hard disks and in DRAM, are there similar studies about how often CPUs hickup? I don't know of any pub
by cperciva 17y ago
There have been many studies that look at error rates on hard disks and in DRAM, are there similar studies about how often CPUs hickup?
I don't know of any published results, but anecdotal reports I've heard from the supercomputing community are in the range of one error per 10-100 CPU years on well-maintained (adequate cooling, clean power, and not overclocked) COTS systems. This rate would be much higher were it not for the fact that most die area is used by caches with internal ECC; I believe that some recent CPUs apply ECC to some internal busses as well.
That said, internal CPU bit errors are a problem which must be considered in high-reliability environments -- this is one of the major factors behind interest in byzantine algorithms.