🚀 go-pugleaf

In article <1gske7INNgje@coli-gate.coli.uni-sb.de> chbl@sbustd.rz.uni-sb.de (Christian Blum) writes:
>
>No way this will work. The refresh rate is set to 18.2 per second; since
>there are 512 rows to refresh, this means every bit in your computer's
>DRAM is refreshed every 20 seconds! There are DRAMS that survive 15ms
>without being refreshed, but a 70ns DRAM can't take it. Try about 10ms
>......

It is surprising how long some DRAMs can go without refresh.  For
production systems, you must believe the manufacturers worst-case
specs, but DRAMs rarely need to be refreshed that often.

A while back, I was working on the software for a computer system that
did its DRAM refresh with a software interrupt -- it only took about
5% of the CPU time.  At first, we had a bug in our refresh code, but
the system was still working remarkably well.  Tracing down another bug
revealed that we weren't refreshing at all, except for the natural
refreshing that came from the execution of code.

Once we got the refresh working, we decided to determine exactly
how slow we could refresh and still have the system work.  We made
sure that the CPU was executing in a tight loop the rest of the time,
so that normal execution wasn't refreshing.

Turns out the particular DRAMs we were using, at the temperatures
we were running, held their contents just fine with a 1-2 second
refresh rate.  They started failing a bit at 5 sec, and were
still 80% correct after 15 sec.

We expanded our experiments a bit, and tried DRAMs from different
manufacturers, and also tried different temperatures.  There
were remarkable differences with both parameters.  We ended up
taking the full 5% CPU hit requirred to meet the manufacturers'
specs, which I think is a good decision for anybody who is tempted
to play around with refresh rates.