How to model cosmic ray corruption of DRAM memory?

[Bob]

I'd like to study how well Chipkill advanced ECC detects
multiple errors in a DRAM memory module. I've written software
that implements the Chipkill algorithm, and now I'm seeking more
detail about exactly how cosmic rays can corrupt DRAM bits.
Such as, when a cosmic ray changes a bit in DRAM, how many
adjacent bits on the average are likely to be affected as well?
What would be the probability density of the number of bits altered
per cosmic ray strike? Of course, it's likely to depend on the
density
of the DRAM. I'd much appreciate any information anyone can
provide on this.

--Bob Day
http://bobday.vze.com

 

[Paul]

I'd like to study how well Chipkill advanced ECC detects
multiple errors in a DRAM memory module. I've written software
that implements the Chipkill algorithm, and now I'm seeking more
detail about exactly how cosmic rays can corrupt DRAM bits.
Such as, when a cosmic ray changes a bit in DRAM, how many
adjacent bits on the average are likely to be affected as well?
What would be the probability density of the number of bits altered
per cosmic ray strike? Of course, it's likely to depend on the
density
of the DRAM. I'd much appreciate any information anyone can
provide on this.

--Bob Day
http://bobday.vze.com

http://www.research.ibm.com/journal/rd/523/dell.html

"Because most upsets affect only a single cell"

the author of the article concludes the Chipkill
covers that pretty well, and has to postulate
a combined hard error plus a soft error, as a
means of defeating Chipkill. (Something like
a four bit hard error, like a dead chip, plus
a single bit soft error.)

In Table 1, the MTTR is measured in months. Would
someone leave a detected memory hard problem in place
that long ?

Paul