[adrian_b]

A buffer is much too small to catch any errors, unless you dedicate more than half of your installed memory to it.

Typical memory rates for a computer with only 8 to 32 GB of DRAM might be of at most a few errors per year when new. For some aged memory modules, after several years of use, the error rate can increase a lot and it can become noticeable. For myself this has been the main benefit of ECC, the ability to detect early the memory modules that must be replaced to avoid data corruption.

Besides the errors from cosmic radiation, which depend mainly on the altitude of the location, there are also errors caused by electrical noise from the environment. The latter may be more frequent in data centers and in industrial computers.

[tzs]

I found my code, and it looks like I was using 128 MB. This was around 2005.

A study done [1] at Google based on 2.5 years of observations of their servers, first published in 2009, found error rates from 25000-70000 errors per billion device hours per Mb. That would put the time Google was doing the study in the right era for when I did my test.

Based on that I would expect to see about 1 error per 40 hours in a 128 MB buffer if the buffer was on a Google server. I suppose that supports the theory that the data center environment is a lot more error prone than my home environment was.

[1] https://cacm.acm.org/research/dram-errors-in-the-wild/