DRAM is susceptible to random occurrences of bit(s) flipping, or changing logical states
unexpectedly. While originally attributed to alpha particles from the device packaging, later research
has shown this to largely be the result of background radiation. There are various influences which
can vary the frequency of the errors, for example chip design techniques, memory density, etc.
DRAM manufacturers strive to accurately model and simulate the soft error rate (SER) of their
technology, but this is a difficult process and is arguably inaccurate. There are also many factors that
affect all error rates in real world scenarios. The obvious system-level challenges exist, such as
ambient temperatures inside an enclosure, contaminants, connector degradation, etc. Additionally,
altitude plays a significant role in the SER. A typical SER might be one bitflip per 2-4 weeks per
gigabyte of DRAM. To put this another way, a 4GB system can expect to encounter roughly one error
each week.
The impact of bitflips to system integrity can vary. On the very innocuous end of the spectrum, the
error can occur in an unused area of memory or simply change the tone of one pixel of color on a
display. Moving toward the damaging end of the scale, a bitflip might cause a program to crash
(typically shut down by the operating system). An error may occur in sensitive data, and if undetected
it may eventually be written to disk and perpetuated. Of course, a bitflip may occur in the operating
system’s kernel, or elsewhere, which can cause an entire system crash. The system may potentially
come down in such a way that the exception handler(s) are not capable of saving the important data
still in memory.