Prime95 / mprime and AVX-512?

[Zarathustra[H]]

So,

I've been stability testing a new (to me) Rocket Lake Xeon E-2314 to make sure it is running properly, before I bring it online.

I'm using mprime (Prime95 for Linux) and I thought it was very weird that while its base clock is 2.8ghz and boost clocks are 4.5ghz, it was humming along at just 3.5Ghz.

Some troubleshooting later, I think I have figured out that this is due to AVX-512.

Since Version 29.8, Prime95 apparently supports AVX-512. and I suspect this is causing all of my cores to run at a lower clock than they otherwise would because of the higher power draw that AVX-512 results in, making them power limited.

I'm not here to discuss the merits of AVX-512 (or lack thereof) but what I am curious about is, will this still serve as a good stability test?

If Prime95 is now using AVX-512, and thus dropping my cores down by a full Ghz, are they still seeing a relatively worst case test, or are they just kind of coasting from being power limited?

If my real life use case doesn't really involve much AVX-512 use, would it be better to download an older version of Prime95 from the archives that predates AVX-512 for proper stress testing?

The general rule when it comes to testing, is tot est as closely to your actual worst case scenarios as possible. AVX-512 may result in worst case power draw, but it certainly doesn't result in worst case clocks.

This may not be a new topic of conversation, I just haven't had an AVX-512 compatible CPU before, so I never had to think about it.

Appreciate any thoughts anyone might have on this.

 

[pendragon1]

look for version 26.6

 

[Zarathustra[H]]

look for version 26.6

Yeah, I already went through the revision history on Prime95 looking for mentions of AVX-512, so I know what version I need to grab in order to make it happen, I guess I just wonder if it makes sense to do so? Would getting an older revision be a better stability test?

It sure would better match the use conditions, that is for sure. Very few applications actually make use of AVX-512.

 

[kalston]

Ideally you want to test as many different scenarios as possible. Full load, medium load, low load, idle... and rinse and repeat with AVX on and off.

But there is no need for an older version? It's just a toggle in recent versions (see at the bottom):

They call it weaker but it stresses the CPU in a different way and with higher clocks so while it may stress the cooler less, it sure is still a worthwhile stability test.

edit: I'm on Windows though, no idea if Linux UI looks like that

 

[uOpt]

I would run without AVX-512. All the history of proving that mprime is a good test has been drawn without.

 

[Zarathustra[H]]

Ideally you want to test as many different scenarios as possible. Full load, medium load, low load, idle... and rinse and repeat with AVX on and off.

But there is no need for an older version? It's just a toggle in recent versions (see at the bottom):
View attachment 631373

They call it weaker but it stresses the CPU in a different way and with higher clocks so while it may stress the cooler less, it sure is still a worthwhile stability test.

Thank you. I had seen the "run a weaker test" option before, but I never knew what it was. I just answered No. Now that I know that it has something to do with turning off and on AVX-512, I am going to have to test it. Thanks for that.

edit: I'm on Windows though, no idea if Linux UI looks like that

It looks a little different, but it looks like it does the same thing.

Essentially it is just the text console version.

Appreciate all the input.

I would run without AVX-512. All the history of proving that mprime is a good test has been drawn without.

That's what I was thinking. I appreciate having a second opinion on this topic. Thank you.

 

[Zarathustra[H]]

I would run without AVX-512. All the history of proving that mprime is a good test has been drawn without.

That's what I was thinking. I appreciate having a second opinion on this topic. Thank you.

Now I have to decide if I should keep AVX and AVX2 on, or turn them off as well.

Ideally there would be one broadly representative test. It sucks to have to run multiple different variants.

 

[bluestang]

Didn't Intel start to down clock the Turbo speeds on v4 and later Xeons when there is an AVX load? So the published Turbo clocks are for non-AVX workloads only.

 

[Grebuloner]

Didn't Intel start to down clock the Turbo speeds on v4 and later Xeons when there is an AVX load? So the published Turbo clocks are for non-AVX workloads only.

Started with v2 by just 1 or 2 bins. The power draw is still just as high as full turbo/TDP, but the voltage is lower and a much higher percentage of the silicon is being utilized.

 

[kirbyrj]

Didn't Intel start to down clock the Turbo speeds on v4 and later Xeons when there is an AVX load? So the published Turbo clocks are for non-AVX workloads only.

They do the same thing with the regular desktop CPUs (although you can set the AVX multiplier yourself with the K CPUs). The only issue is that I don't know if it is regularly advertised anywhere what it is.

 

[xDiVolatilX]

Lately I've discovered Ycruncher. While I'll admit I don't understand shit from what I'm looking at, Buildzoid always uses it so I ran it a few times and passed. Looking into learning more about it.
What I've noticed it Ycruncher pegs it at 100% more consistently than prime, not sure why because Ycruncher is supposedly more of a brutal stability test.

 

[Zarathustra[H]]

I have now tested with and without AVX-512 on the E-2314, and to my surprise all core turbo maxes out at the same 3500Mhz regardless of whether AVX is enabled or not.

It will hit 4500 Mhz (max turbo) in single core with AVX-512 off.

In single core with AVX - however - it does reduce the clock slightly to 4400Mhz.