I am running a Ryzen 7900x with a custom loop which consists of a 120mm and a 240mm rad.
I have been experimenting with a manual overclock and testing with prime95. thoughts
A lot of misinformation about prime95
not a lot of documentation on the behavior of CPU/ cache while running prime95.
Questions
So while running prime 95 small FFTs with FFT in place selected, I have noticed that once the FFT gets larger than 200k that my core utilization will drop to about 99.7 percent utilization on all threads. Is this because the FFT is too large to fit in the CPU cache and that it is leaking out into memory?
Really, you want to know where exactly the 0.3% went?
This level of curiosity is certainly admirable.
You can start by disabling hyperthreading; it does not do anything useful for LL tests.
Also, if you are running into thermal throttling anyway (I do on a 7950x all the time, with a 420 AIO), then anything else does not really matter, does it?
Third, the time per iteration should be a more usefull metric than the CPU utilization percentage; me thinks…
Good luck with your investigation. It surely is interesting.
Hello, thanks for the response, Yes a very curious individual indeed. I want to understand, I have to understand < Oo > But yes I have a hypotheses that FFT over the length of 200k are too large to fit in the cache of the processor including L1/L2/ and L3 therefore the processor utilization will drop slightly because the FFT will spill over to memory then the cpu first has to retrieve the data from memory then complete the calculation which will be slightly slower than storing the data in cache and completing the calculation. which would result in a core and thread percentage usage drop and slightly lower effective clocks.
@TObject and @vic so looks like I was a bit off on the size of the FFT looks any FFT over 240k and processor utilization drops off a fraction. I am making some inquires at the Mersenne forums to do some additional research I will get back with my findings.
Prime95 forum is perhaps a much better place to seek feedback on P95. The author or current maintainer is often there last time I checked.
I didn’t pay much attention to CPU utilisation. I did see power consumption is lower in VM when running the same test. AMD Ryzen seems to have quite a few weird things that AMD perhaps will never tell the public. And I’m zen2. I believe later gens are likely improved. AMD internal validation should have started including Prime95, either since zen3 or if not surely zen4.
I’m not into prime numbers but Prime95 is a terrific tool for testing CPU & memory stability. Keep us posted.