It helps to use a more stressful test, than a light computing load.
For example, memtest86+ from
memtest.org, can be used to test the
memory. Usually, this is available in floppy and also in ISO9660
for a CD. The test will run forever, unless you stop it. No errors
are acceptable. I've had two batches of cheap RAM, that started
throwing errors after the warranty on the RAM was expired. The
failed locations in the RAM were consistent from one run of
memtest86+ to the next. (It is also possible for the errors
to be random, when there are memory bus problems.) But the
important message here, is that no errors be reported.
Microsoft also makes a memory tester, free for download.
If you pass that test, then getting a copy of Prime95 from
mersenne.org
is a second test. It runs in Windows or in Linux. There are a few
different versions (including one that spawns a test thread per
core on multicore processors). It uses hand coded FFT (fast fourier
transform) code, where the author of the code knows what the answer
of the math calculation should be. It is a pretty sensitive test of
not only the memory, but also the processor and Northbridge.
This version is suitable for even quad core processors, and is a
Windows version. Select the stress test option (rather than joining
GIMPS), and accept the default test conditions for your first test
run. On a mis-adjusted computer, this program will report a problem
in as little as ten seconds. I use a four hour run, error free, as
my current success criterion.
http://www.mersenne.org/gimps/p95v255a.zip
Sometimes memtest86+ passes, and Prime95 fails. Prime95 tends to be
a more sensitive test, mainly because the computer gets so much
warmer, and there is also more electrical noise from the test. But
you really shouldn't be booting into Windows, if the memory is
extremely bad, which is why testing first with memtest86+ is
recommended as a first step after a memory changeout.
Paul
