Do I need to run 2048K+ in Prime95 to stress test my XMP profile?

steffeeh · May 26, 2019

I've built a PC for a friend, and I'm currently running some tests to make sure everything work as it should.

If I've understood things right, normally when you overclock the RAM you want to run Memtest86, and then Prime95 in custom mode with a range of 512K-4096K, where 512K-1024K stresses the IMC and I/O-lanes, and 2048K+ stresses the actual DIMMs (as outlined by the RAM overclocking guide on this forum).
However now I only want to see if the XMP/DOCP is stable and works well in the system - do I then really need to stress test the DIMMs themselves with 2048K+ FFTs? Would be enough to just stress test the 512K-1024K range?
They are after all binned, and should work running at the marketed specs assuming the rest of the system works with the profile too.

Truth be told, he's picking the system up tonight so I'm a little short on time and can't run Prime95 for 8+ hours without telling him to come by tomorrow instead.

jj9987 · May 26, 2019

If Memtest86 didn't show any issues and the system is stable, I wouldn't bother tbh. Memtest is already pretty thorough.

scottyseng · May 26, 2019

memtest would be the most ideal case for this. Otherwise if you're short on time, run Intel Burn Test on the maximum preset (will take all RAM). If it survives 5 passes, it's generally really stable. I usually do 10. It does take longer to run depending on how much RAM you have.

I've yet to see an XMP profile be unstable though (for reasonable speeds)

porina · May 26, 2019

memtest more tests the ram than the IMC. Prime95 does push the IMC more too, I've seen that fail where memtest will run as long as you like without error. To keep it simple, I'd just do a large FFT test and leave it as long as you like.

For a more advanced consideration, the calculation ram per worker is roughly 8x the FFT size. e.g. if you use 512k FFTs, it will take 4MB of data for that one test. Then you need to multiply it by the number of threads running. You can then compare this value to the processor cache. If you're much above the L3 total quantity, you're going to be hitting ram a lot.