Question Need help understanding CPU stats during Prime95 runs on 7950x

[keithth]

Despite watching some youtube videos, I'm new to water cooling, this new AMD platform, and overclocking in general.

I was seeing 95C+ temps during Prime95 (small FFTs) NH-D15 7mm offset runs, which brought my median average effective clock to 4687.

Didn't like this result, so I installed an Arctic Liquid Freezer II 360 aio, and now I'm seeing temps around 80C during the same test, with an average effective clock of 4775. Better, yeah, but not great.

I don't understand why the 7950 isn't taking advantage of the available thermal headroom, and running faster. My theory is that the because the PPT is coming close to maximum (around 96% of 230W), that this is artificially limiting clock.

Does this sound like what's happening and are there some minor, safe, tweaks I can do to increase the performance?

Thanks

 

[Darkbreeze]

What is the model of your motherboard?

What is the model of your case?

What does your case fan configuration look like, as in, how many case fans are installed, where are each of them installed and exactly what direction is each of them oriented to, intake or exhaust?

This sounds to me like a problem, not with core temperatures, but with VRM temperatures causing thermal throttling.

What are you using to monitor thermals and other sensors? If you are not using HWinfo, I would highly recommend that you download it, install it, run it, choose "Sensors only" and take a look at all the sensors including those for the motherboard and VRM temperatures and check to see if thermal throttling is indicated to the right of the field that indicates thermal throttling which, however, only indicates if there is throttling due to core temperatures. You'll have to watch the VRM MOS, System and PCH, or whatever is listed in the section for your board sensors, while you start and run Prime95 Small FFT. Be sure you do not choose "Smallest FFT" and that you have disabled ALL AVX options. Some options may remain gray until you disable other options but in the end all AVX options should be disabled while doing thermal testing.

 

[keithth]

Hello again! I think you helped me spec this system originally!

The current config is here: https://pcpartpicker.com/list/kxMQKX

Regarding case fans, I've removed the previous front 140mm's I had in place, to make room for the (3) 120mm's of the AIO. The only other real fan (besides the tiny VRM fan on the arctic cpu head) is the 120mm exhaust in the rear.

I've been using HWinfo to collect the sensors only data, creating .csv's for the results.

The VRM MOS ramps from 33C at the beginning of the test and then maxes out at 51C. The thermal throttling columns are all essentially NO's. PCH shows 46C-50C. System 27C-30C.

Thanks!

 

[TerryLaze]

Prime 95 is a stress test and will have lower clocks than pretty much anything else.
You have results pretty close to this one with PBO enabled, if you want to you can use the curve optimizer as shown there to get a bit more clocks and reduce the voltages.
Probably not worth all the trouble for a normal user since it's pretty involved.

SkatterBencher #45: AMD Ryzen 9 7950X Overclocked to 5950 MHz - SkatterBencher

We overclock the AMD Ryzen 9 7950X up to 5950 MHz with using the GIGABYTE X670E Aorus Master motherboard and EK-Quantum water cooling.

skatterbencher.com

 

[Darkbreeze]

So you have three intake fans, but they are on the AIO cooler in front, and one single exhaust fan, 120mm in the rear? Is that right? And you still have the two 140mm fans that were in the front?

While I don't actually think with those temps you're truly seeing a thermal problem, it's also quite possible sometimes that a given temp might LOOK artificially low because it's already throttling and HWinfo is only going to tell you if you are seeing throttling from core or package/ring temperatures, at least on Intel platforms. I'd have to go out to the garage and take a look at my Ryzen system in HWinfo to see if there are other throttling indicators on AMD chipsets but I don't think there are.

Anyhow, in my opinion I'd take those 140mm fans you removed from the front and I'd stick them in the top-rear and top-middle locations as additional exhaust fans. This will serve several purposes.

1. It will help to provide some additional airflow over the top of your VRMs, so that IF there is, now or later, any issues with them getting warm it should help to offer at least some assistance in keeping them cooler.

2. It will reduce the internal case pressure, which will help boost internal cooling for everything by both getting some additional heat that tends to congregate at the top of the case, out faster, plus the reduction in internal case pressure will make it moderately easier for your front intake fans to operate since they already have to struggle against the static pressure of the radiator so anything you can do to make things easier on those fans means they will be able to work a bit more efficiently. Plus, since you've got an AIO there you really want any heat that is being shed from that radiator to get out of the case as fast as possible and those additional exhaust fans will help to expedite that process as well.

3. With two 140mm exhaust fans and a 120mm exhaust fan, instead of just a 120mm exhaust fan, you SHOULD be able to somewhat reduce the speed that the exhaust fans need to run at while still getting more heat out of the case than you ever could using just the single exhaust fan, which should help to reduce noise levels. Three low to mid speed fans are going to always be quieter than one fan running for it's life under higher loads.

And you might also want to consider custom configuring the curves for the exhaust fans and the AIO in your BIOS. If you haven't already done that and wish to do so I can provide you with a screenshot of what the case and CPU cooler fan curves look like in my system and then you can play around with it until you get where you are getting good cooling but also generally low noise levels, if you don't already have it set up where you are good with it.

Aside from that, another VERY important consideration that I asked previously but that you did not answer, is what your current motherboard BIOS version is? This is VERY important as many bugs or component behaviors get tweaked by the manufacturer pretty often along the way and it's not like the old days where you should only update if you absolutely have to. It's a bit more like driver updates these days, where it's more often a necessary process periodically.

 

[mdd1963]

The function of PBO is to extract everything from your CPU that it can just short of throttling....so unless you seeing abnormally low clocks across all cores, I'd not worry about high temps in Prime95 based on that alone, as long as your temps are 70-80C or less in normal workloads, like...gaming!

Prime95 is quite stressful, especially the classic Small FFTs with AVX/AVX2 enabled; the old defacto standard was to do small FFTs with AVX/AVX2 disabled, as this was considered a '100% load of CPU', where as AVX enabled was actually an 110-115% overload according to some well versed CPU engineers...

 

[Darkbreeze]

Right. But the problem here is that his maximum temperatures are now about 15°C below throttle temp spec, which is good, but he is not hitting anywhere near the expected ~5.1Ghz all core that he should be given the amount of breathing room still present in his thermal envelope.

Since we know that this CPU will not achieve even single core 5.85Ghz when it is over 50°C I suspect that maybe there are similar conditions in play but it really makes no sense that it would be limited to 4.7Ghz at 80°C as there is still plenty of headroom available before any throttling should occur unless it is already occurring, which is why I wanted to verify that while running Prime95 they are in fact running Small FFT, and not Smallest FFT, and that ALL AVX options are disabled during the test. Also that they are running the latest version of Prime95.

Another option would be to run OCCT CPU stress test with settings "small data set", "normal" mode, "steady" load type and "SSE" instruction set, to see if you are seeing identical behavior as with the Prime95 testing. Likely, it will be the same but it's worth looking at anyhow.

 

[mdd1963]

Not sure if AMD mainboards have the AVX offset values in BIOS like Intel boards? (this will, if/when AVX offset enabled, intentionally hit lower clocks by 100-400 MHz for values of -1 thru -4, when encountering an AVX workload, which Prime95 does enable by default.....)

Might be worth checking....

(Edit: sorry, I see the AVX issue already mentioned above! Doh!)

 

[keithth]

Thanks @mdd1963, @Darkbreeze, and @TerryLaze for the discussion.

A few things, I'm using Prime95 from guru3d which is version 30.9 build 3. If the official source is https://www.mersenne.org/download/ then 30.8b17 must be the latest?? I was not disabling AVX, which there are 3 or 4 related AVX checkboxes. So any results are with the default. I was using SMALL FFTs, not smallest.

While I must have missed the original request for the bios firmware version, I am using the latest available on Gigabytes site -- F20a from November 10th. This definitely fixed stability issues with my memory by upgrading to it. I upgraded day 1, so any quoted results already have this applied.

These 140mm noctuas, that I still have, but not currently installed are still pretty quiet. I can play with the fan curves..... I don't mind some fan noise during intensive tasks.

I've modified my case to allow me to go tubes down with the radiator in the front. The power supply shroud was getting in the way. But it ain't no mo'. I've got some u-channel rubber coming in the mail soon, and I'm going to rearrange the fan/radiator configuration. And I can add those 140mm fans --- don't have a current use for them anyways. As exhaust.

Right. But the problem here is that his maximum temperatures are now about 15°C below throttle temp spec, which is good, but he is not hitting anywhere near the expected ~5.1Ghz all core that he should be given the amount of breathing room still present in his thermal envelope.

Precisely! I'm not having an (obvious) CPU thermal problem, rather, I want to push the CPU harder because I've got the headroom to do it! It's not clear yet to me why the CPU isn't automatically increasing clock speed!

Let me get the GIMPS version, make sure AVX disabled, re-run my test to see if I'm seeing similar results. I'll also run that OCCT.

Thanks!

 

[Darkbreeze]

Guru3d must be full of crap, and in light of that I wouldn't even trust the download from them at all because there ISN'T any 30.9 version. Get it directly from mersenne.org because they are the ones who MAKE the utility in the first place, so whatever shows as the latest version there IS the latest version and that is where I'd get it from and what I would use.

Tubes down is correct.

I suspect that when you run it with AVX disabled, and you will first need to uncheck one version of AVX before it will allow you to uncheck additional versions in my experience, you will likely see a couple hundred Mhz faster CPU operation. But yes, also try OCCT if there is no change, or even if there is. About five minutes of either test should be all that's needed to get a good idea of what the all core max out is going to be. You might also download and run Cinebench since it's not an all core steady state test utility and see what your core usage looks like while running that as well because it's probably going to be more representative of a real world intense workload. You still want to know that you are able to maintain thermal compliance with Prime95 Small FFT and AVX disabled for like ten minutes, but that's a different thing, thermal compliance testing, than what we're trying to discover right now. Yet, still a thing you'll probably want to do just to be sure.

Since you are not manually overclocking I don't expect a major issue in that regard but if you are running PBO then it's a good idea to do it anyway because it IS still an overclock, it's just an automatic one that is system controlled.

You might even want to disable PBO in the BIOS after you do the initial testing if nothing changes, and then re-test with it disabled to see if you are able to maintain higher all core clocks while testing because I have seen some systems actually get BETTER performance without PBO enabled on some CPU models simply due to being able to maintain higher all core clocks while under a full load. It's not typical, but it is possible.

 

[keithth]

Ok here's some visuals of results. Looking much better up over 5000. OCCT makes the CPU run hotter.

I changed to the official version of Prime95, disable AVX (Which did require 3 different disables), and updated hwinfo to latest (it was two minor revs behind).

My latest Cinebench 24 scores are multi-core 2200, single-core 125.

I'm going to add those 140mm in top because they are just sitting here now.

Thanks
https://i.imgur.com/rgZ5sD3.jpg

 

[Darkbreeze]

I think your numbers are wrong on the Cinebench scores, because average scores for the 7950x indicate that single core should be somewhere around 2205 and multithreaded should be around 29,650 or in that range anyhow. Definitely cooling plays a role as does CPU sample quality.

But otherwise, seeing that you are indeed over the 5Ghz all core threshold is definitely another improvement and is about where you ought to be.

 

[keithth]

Are you looking at Cinebench R23? Or Cinebench 2024?

 

[Darkbreeze]

Yeah, R23. For R24 your numbers are actually about exactly where the averages fall so I think you are in good shape.

 

[keithth]

Thanks. I just found an older copy of Cinebench R23, which appears to no longer be downloadable directly from Maxon.

I'm posting 37936 multi and 2018 single in R23.

 

[Darkbreeze]

Those seem good enough too. Obviously you could push things with a manual overclock, but ever since the 3rd Gen Ryzens came out, manual overclocking on Ryzen seems like a waste because they are already pushing the envelope with the PBO configurations and they just "work" at least as good or better in most cases than a manual overclock. I'm sure there are exceptions and with the cooling you have you might be able to get away with slightly better all core manual clock settings, but it probably won't be worth the amount of effort you'll have to put into it.

I'd install those 140mm fans and see how that affects things.

 

[keithth]

I've got some more testing to do, but I might be going backwards a bit.

I installed my AIO in a non-ideal, but supported, configuration originally, although tubes down like I said, where the radiator was in front of the (3) fans. Despite having built a fair number of machines, I've never done AIO or RGB. I initially didn't hook up the RGB, but once I did, realized my error because I couldn't see the RGB lighting from outside the case through the front. Maybe it's a stylistic choice. So the air was being pulled through radiator.

Anyways, I've swapped them around, which was kind of a royal pita, with the fans attached to the front of the case, and radiator behind them. Installed the 140mm fans up top as exhaust.

Numbers are back down into the 4600-4900mhz region.

I'm thinking of putting a fan on the inside top attached to the back of the radiator.

 

[Darkbreeze]

The problem is, RGB and ARGB fans generally don't have very good static pressure. High static pressure ratings are very much desirable on any fan that has to overcome the restrictions of a radiator or heatsink, or when in an intake configuration. Adding those 140mm fans to the top of the case, in the top-rear and top-middle locations, as EXHAUST, blowing OUT of the case, should help reduce some of the internal pressure so that the front fans bringing air in through the radiator don't have to work AS hard, but the fact is, with them being the RGB variant, they are not as good as the Arctic fans that come on the Liquid freezer II coolers that are not RGB.

I'd say screw the RGB and put three good high static pressure fans on the radiator like the Noctua NF-F12 PWM chromax.black.swap fans or something with a similarly high static pressure. You'll be hard pressed to find another fan with the same kind of static pressure as the Noctua NF-F12 PWM or NF-A14 PWM chromax.black.swap fans. Or even just the regular Noctua NF-F12 PWM or NF-A14 PWM fans which are the Noctua "Earth tones", better known as brown and tan.

But, if it was getting better temps with the radiator fans in pull configuration, then moving them back to that is fine too. The static pressure limitation is still applicable though, whether in push or pull on a radiator or heatsink.

 

[keithth]

Appreciate the advice. I ordered (3) more Noctua 12 blacks, should be here today. While I was briefly excited to have some RGB, I won't trade performance for the looks. Once I get this running right, then I'll go back and add some bling.

I would not have guessed that the argb fans were worse than the regular ones. I figured they just tacked on some lights on the normal fans. I will say that the fans felt cheap, and noticeably lower build quality than the dozens of Noctua's I've bought over the years.

Despite owning a bunch of them, those brown and tans are just plain ugly.

I will say one of my least favorite things of this build has been messing with the AIO. Too many screws. Large radiator. Tight clearances in the case. Tubes just a bit shorter than I'd prefer.

 

[keithth]

Got those (3) noctua 120's installed replacing argb arctic's, and repasted with kyronaut.

So current config has:

(1) 120mm exhaust back
(2) noctua 140mm exhaust top
(3) noctua 120mm intakes on outside of the radiator
(1) 120mm intake on the top inside of the radiator

Fans manually set on maximum.

18.6C ambient

Idle temps are around 38C.

Average CPU temp was about 82C, only peaks during Prime95 noavx at 86C for a very short period.
Average Effective Clock was 4876, peaking at 4951.

I'm not sure I can throw much more cooling at this.

And besides..... these temperatures do not appear to be the problem..... I'm so far away from 95C.

 

[Darkbreeze]

I would take that 120mm fan off the inside of the radiator unless it is the same model as the three Noctua's you put on the outside of the radiator. In reality, it is probably causing more harm than good because if it does not have the same static pressure and CFM capabilities that the three fans that are pushing have, then it is simply becoming another obstacle that the top fan has to overcome to push any air through the radiator. If you are going to use a push pull configuration, on ANYTHING, you want all the fans to have the same or VERY similar static pressure and CFM capabilities.

But, I think I'd have to agree that at this point it wouldn't seem like cooling is the problem. Especially not CPU cooling. And with those Noctua's up top you should be getting plenty of residual airflow over the VRMs besides which since I still haven't seen any HWinfo sensor values while running prime. What utility are your previous screenshots from?

I would download and install HWinfo, run it, choose "Sensors only" and then run Prime again while HWinfo is running. Keep HWinfo running while Prime is running, for like five minutes or so, then take screenshots with Prime still running, of ALL the HWinfo sensors. You will probably need to take several screenshots and scroll down between screenshots to capture all the sensor values. What I'm primarily interested in are the motherboard 1, motherboard 2, VRM and other board related sensor values, in addition to the throttle information and CPU values.

Where is your pump connected? Is the header the pump is connected to configured to run at 100% speed at all times? Because it should be if it's running off one of the motherboard fan headers. Pump should not be variable speed except on models that don't give you any option. If your board has an AIO header, you should be using that for the pump and it should be set to 100% full time. Not sure exactly what you configuration has because Arctic has changed the configuration between versions a few times.

 

[keithth]

Ok, a few things... rearranged the fans like I had them originally in a pull (so inside, fans on intake, radiator, outside) configuration, because that's where I saw the best temperatures.

I've been using HWinfo with sensors only logging the whole time, while running Prime95. The screenshots you see are from the Hwinfo-provided "GenericLogViewer." HWinfo saves extremely wide .CSVs, taking snapshots every 2 seconds for the entire period of the run. So one row generated per sample. There's no real need to screenshot. Every sensor displayed on screen is included in the .csv's.

The pump and the radiator fans are not separate connections. There's one singular 4-pin connection which I have connected to the CPU_FAN header. Well, and the ARGB cable which I'm not using.

https://www.arctic.de/us/Liquid-Freezer-II-360-A-RGB/ACFRE00101A

I could alter this configuration. I don't HAVE to connect the fans to the supplied cable. I could leave the pump connected as normal, and then separately wire the fans to a different header. It's occured to me, though, that maybe the design calls for variable speed fans with a (matching?) variable speed pump. That's the default recommended hookup method, after all.

I enabled auto overclocking, and the maximum speed displayed in Ryzen Master goes from 5800 to 5950. And then further Prime95 runs at just about 5.0ghz. And course, a corresponding temperature increase to 95C. So that behavior seems to what I want.

I can get a .CSV uploaded and made available for you, if you're interested.

I'm starting to reach a saturation point of messing around with this..... and it's starting to feel overly academic to me....... I may just leave the OC alone, and be happy with the 4.9ghz. It's crossed my mind to order a suitably sized case and 420mm AIO, but maybe I'll leave that for another project when I'm bored.

This article is the thing that chaps my butt.

https://www.techpowerup.com/review/amd-ryzen-9-7950x-cooling-requirements-thermal-throttling/

The talk about getting 5171mhz with a SINGLE stack Noctua Cooler..... which I couldn't get with the D15 with both fans running at 100%. And also can't get even with 360mm AIO.

Thanks.

 

[keithth]

Ok, Please see attached image. Tried to include the relevant items.

The HWinfo64.exe logging output, set to an increased frequency of one sample every 1/2s.

I'd suggest downloading the log viewer to make it easier to digest. This is not really human readable in this form....

https://pastebin.com/5GiEPZFP

Notice that the three numbers are minimum, average, maximum.

Thanks

 

[keithth]

Just another data point, that without OC enabled, I'm seeing 5.2ghz on 8-thread Prime95 runs.