[SOLVED] Weird Ryzen 3700X temperatures

Sergei Tachenov

Commendable
Jan 22, 2021
64
64
1,610
I'd hate to start yet another “My Ryzen is running too hot!” thread, but this is a little bit different...

Setup: Fractal Design Define S2 with three stock GP-14 fans, top panel closed, neat cable management, Ryzen 3700X with Arctic Freezer 34 CO, all fans connected to the case hub which is connected to CPU_FAN at MB and set to 40% PWM which is about 840 RPM for the CPU cooler fan and god knows what for case fans (only one fan is reported, obviously), but the overall setup is dead silent. No overclocking or any non-standard setup except that I have disabled Cool'n'Quiet because my CPU is normally under heavy load, so it probably wouldn't do any good.

Running BOINC (World Community Grid) 24/7 at 60%, I get about 75C (temperature inside the case reported by MB is 27–28C). Going up to 80% I get steady 80C. Going up to 95% I get up to 85 or something. So far it makes sense...

But when I start Prime95, things get real weird. First thing that happens is that temperatures slowly drop to around 70C. Well, my first guess was that Prime95 gives more stable load: 95–100%, while BOINC goes up and down between 0% and 100%, averaging at 60%, as requested. The CPU clock goes haywire, and that could be the reason why it runs hotter than at just steady heavy load.

But then, after running Prime95 for a couple of minutes, temp suddenly jumps up to 80C. Then it is steady again. Then it suddenly drops back to 70C. Then it gradually increases. Upon reaching 75C it skyrockets almost to 90C. Then it gradually decreases (this makes sense as the CPU dropped the clock from 4000+ to 3800+ here), steadying at 80C again. CPU load stays at 95–100% all the time, no other heavy tasks running. Just what on earth is going on here? I'm very curious...

The funniest thing is that if I just increase BOINC usage to 80% or even 100%, none of this happens. Things just get gradually hotter. Must be a Prime95 thing...

The temps:

PiVUFdWe_o.png


The clock:

u684Ubif_o.png


CPU usage in the Task Manager:

fQdnVd8a_o.png
 
Solution
...
Must be a Prime95 thing...
...
I can imagine you're running the default Prime95 torture test setting, which is 'Blend'. In blend mode it cycles through a variety of FFT (Fast Fourier Transforms) sizes and cache utilization combinations which also present an equally varied amount of CPU loading for arcane reasons, known mainly to the coder and advanced mathematicians. When running very large FFT's, and using more of main memory, it's much easier on the CPU while small FFT's and 'in place' it's much harder and so the CPU temp rises a lot higher. The difference in cpu processing load can be HUGE as it changes in 'Blend' mode and that will cause massive temp changes. It's also far more pronounced with Ryzen 3000 than previous...
I'd hate to start yet another “My Ryzen is running too hot!” thread, but this is a little bit different...

Setup: Fractal Design Define S2 with three stock GP-14 fans, top panel closed, neat cable management, Ryzen 3700X with Arctic Freezer 34 CO, all fans connected to the case hub which is connected to CPU_FAN at MB and set to 40% PWM which is about 840 RPM for the CPU cooler fan and god knows what for case fans (only one fan is reported, obviously), but the overall setup is dead silent. No overclocking or any non-standard setup except that I have disabled Cool'n'Quiet because my CPU is normally under heavy load, so it probably wouldn't do any good.

Running BOINC (World Community Grid) 24/7 at 60%, I get about 75C (temperature inside the case reported by MB is 27–28C). Going up to 80% I get steady 80C. Going up to 95% I get up to 85 or something. So far it makes sense...

But when I start Prime95, things get real weird. First thing that happens is that temperatures slowly drop to around 70C. Well, my first guess was that Prime95 gives more stable load: 95–100%, while BOINC goes up and down between 0% and 100%, averaging at 60%, as requested. The CPU clock goes haywire, and that could be the reason why it runs hotter than at just steady heavy load.

But then, after running Prime95 for a couple of minutes, temp suddenly jumps up to 80C. Then it is steady again. Then it suddenly drops back to 70C. Then it gradually increases. Upon reaching 75C it skyrockets almost to 90C. Then it gradually decreases (this makes sense as the CPU dropped the clock from 4000+ to 3800+ here), steadying at 80C again. CPU load stays at 95–100% all the time, no other heavy tasks running. Just what on earth is going on here? I'm very curious...

The funniest thing is that if I just increase BOINC usage to 80% or even 100%, none of this happens. Things just get gradually hotter. Must be a Prime95 thing...

The temps:

PiVUFdWe_o.png


The clock:

u684Ubif_o.png


CPU usage in the Task Manager:

fQdnVd8a_o.png
Those temperature drops are due to CPU starts throttling down (mainly by lowering voltage ) when overloaded. That also triggers frequency drop and so on.
That cooler is just not good enough for such sustained loads, ad to that limited RPM on fans and no wonder it behaves like that.
Even my water cooler is sitting on the edge of 70c going up to 80+ under sustained high loads. Best working temps are between 62-70c above which it starts loosing on boost, about 100MHz for every 2c above 70c.
 
  • Like
Reactions: Krotow
Those temperature drops are due to CPU starts throttling down (mainly by lowering voltage ) when overloaded. That also triggers frequency drop and so on.
That's exactly the source of my confusion. After switching from 60% BOINC to Prime95, the clock is steady about 4200 MHz (look at the 2nd screenshot). It only drops when the temps approached 90C, which is the only part of this that makes sense to me. Or are you saying that throttling actually begins earlier, it's just that the voltage drops and not the clock?

But even then it doesn't explain at all why the temperatures went down for a couple of minutes when I first increased the load, and then literally jumped to 80C. Like, it started throttling and then suddenly decided to ease on it a bit? This PBO looks some heavy A. I. stuff then...

Or maybe it's just that the nature of Prime95 load changes? Like, it is performing some initialization first and then does some heavy work later, and even though both give high load, they somehow use different parts of the chip? Sounds like crazy nonsense to me, but I've no knowledge of how the CPU really works.

Off-topic digression:

High temperatures themselves aren't surprising me... I realize that the cooler is cheap, and that RPM is not something that could lead to efficient cooling. The cooler surprisingly good at max RPM, though, and not that loud in this mode. Steady fan hum at 1800 RPM gives me about 75–77C under 100% BOINC (and different weird changes with Prime95), which looks nice for a budget cooler.

I only bought this cooler because I got confused by 65W TDP, thought that any cooler would do just fine, compared to my old FX-8320 @ 125W. Didn't take high density into account. I'm even willing to buy a better one, but I'm having a hard time finding one that is good enough at low RPM. Most tests and reviews seem to run at 1000 RPM, which is a bit too much for me. In fact, even 1800 is better for me than 1000, because I am experiencing very annoying resonating noises in the 950–1400 range.
 
That's exactly the source of my confusion. After switching from 60% BOINC to Prime95, the clock is steady about 4200 MHz (look at the 2nd screenshot). It only drops when the temps approached 90C, which is the only part of this that makes sense to me. Or are you saying that throttling actually begins earlier, it's just that the voltage drops and not the clock?

But even then it doesn't explain at all why the temperatures went down for a couple of minutes when I first increased the load, and then literally jumped to 80C. Like, it started throttling and then suddenly decided to ease on it a bit? This PBO looks some heavy A. I. stuff then...

Or maybe it's just that the nature of Prime95 load changes? Like, it is performing some initialization first and then does some heavy work later, and even though both give high load, they somehow use different parts of the chip? Sounds like crazy nonsense to me, but I've no knowledge of how the CPU really works.

Off-topic digression:

High temperatures themselves aren't surprising me... I realize that the cooler is cheap, and that RPM is not something that could lead to efficient cooling. The cooler surprisingly good at max RPM, though, and not that loud in this mode. Steady fan hum at 1800 RPM gives me about 75–77C under 100% BOINC (and different weird changes with Prime95), which looks nice for a budget cooler.

I only bought this cooler because I got confused by 65W TDP, thought that any cooler would do just fine, compared to my old FX-8320 @ 125W. Didn't take high density into account. I'm even willing to buy a better one, but I'm having a hard time finding one that is good enough at low RPM. Most tests and reviews seem to run at 1000 RPM, which is a bit too much for me. In fact, even 1800 is better for me than 1000, because I am experiencing very annoying resonating noises in the 950–1400 range.
If you are willing to go for an AIO cooler, I'd suggest Arctic Liquid Freezer II at lest like mine 240 or better. Arctic fans like P12 and P14 are among quietest ones, I don't hear then even at full speed (1800rpm+),Ryzen is so sensitive to temperatures that it really responds best to cooling. If you have set OC than temps have much less influence than in auto mode as long as they are under 90c.
 
No AIO, thanks. Just too much trouble. Especially just for BOINC. I'm only gaming at 1920x1080@60, so during gaming temps are even lower than when BOINC runs. But even when I eventually get a better display, I'll probably stick to air, just to avoid worrying about possible leakages.

No OC either. Just PBO with default settings.

The fan is quiet. And it's also supposed to be reliable as hell because of the dual ball bearing. Of all fans I've ever used I probably like this one most. It was a very pleasant surprise for a budget cooler. But I'm somewhat sensitive to noises, so I prefer to keep everything silent, not just quiet.
 
In general I agree with CountMike (above).. just wish to add my thoughts, as I had this CPU.
First, I don't think your cooler is "that bad" for 3700X. I used beQuiet Dark Rock Slim, which seems to be similar performer (but kinda expensive) and I had no issues.
PBO.. I didn't really dig into that, but yes, it can make CPU running quite hot -with not much benefit in performance. I ran few benchmarks and decided to turn PBO off.
About "full load".. it also depends from application to application. Means which "cores" are being used at the same time. Yes, in Windows, you get listed 16 cores, but only 8 of them are "real" cores. And if it happen (only) those real cores are under heavy load, CPU temperature can increase substantially.
You can make an experiment with CineBench (for example). Open CineBench, run it (multicore test) and observe temps. After benchmark ends, keep CineBench open, disable "cores" 8 to 15 (in Windows TaskManager) and repeat bench. You will see, CPU will be hotter by running 8 cores at full load than if running all "16" cores under full load. That's because cores 0-7 are real cores, while upper eight "cores" are threaded (simulated, so to speak).
In short, disable PBO and make sure CPU cooler is mounted properly. Finally, if not sure about cooler, I recommend Scythe Mugen 5 (about 50 US$) which I use on 5900X.
 
Yes, I realize there are 8 physical cores. Windows actually shows both numbers, like “8 cores” and “16 logical processors”.

Interesting. CineBench doesn't give any of that weird Prime95 readings. Just steady temperature corresponding to what I saw with BOINC.

Turning PBO off, however, lead to immediate drop to 45–46C at 1800 RPM and 52–53C at my favorite 840 RPM. Cool and silent! CineBench points dropped from 12600 to 11000, which roughly corresponds to 15% drop from 4200 MHz to 3600 MHz. Makes sense. Unlike Prime95. Guess that was a Prime95 thing after all.

Maybe I should keep it off after all... At least until I actually need that performance. This way I can run BOINC at 95% 24/7 staying below 55C in dead silence.
 
Maybe I should keep it off after all... At least until I actually need that performance. This way I can run BOINC at 95% 24/7 staying below 55C in dead silence.
-yes, I think that way too. Of course, there's performance drop without PBO, but not that much really IMO. I just see PBO as "brute force" (instead of "smart") solution.
I've read interesting article at Guru3D about more "intelligent" overclocking.. -or should I say "optimal tuning". Sure worth to check, I think.
 
Oops, it seems I've mixed up PBO and Core Performance Boost. It's the latter that I've disabled. I never have PBO enabled (I'm not really into OC), while CPB was enabled by default.

The really weird part is that I got two BsODs since I disabled CPB. One was during CineBench benchmark, the other one was when just harmlessly typing something in a browser with 95% BOINC in background. So running at 80–85C for a few days is OK, and 53C at 3600 MHz causes BsOD? Figures.
 
...
Must be a Prime95 thing...
...
I can imagine you're running the default Prime95 torture test setting, which is 'Blend'. In blend mode it cycles through a variety of FFT (Fast Fourier Transforms) sizes and cache utilization combinations which also present an equally varied amount of CPU loading for arcane reasons, known mainly to the coder and advanced mathematicians. When running very large FFT's, and using more of main memory, it's much easier on the CPU while small FFT's and 'in place' it's much harder and so the CPU temp rises a lot higher. The difference in cpu processing load can be HUGE as it changes in 'Blend' mode and that will cause massive temp changes. It's also far more pronounced with Ryzen 3000 than previous AMD processors, at least compared to the FX-6300 and R7 1700 I had previously.

You can force to run in a very heavy load that will always heat the CPU the maximum amount by running Custom torture test settings: Min and Max FFT of 128K, Run FFT's in-place (so it loops out of local cache I guess?). DO NOT disable AVX2(fused multiply-add) for maximum heating, but then disable AVX2 and then AVX to see how significant it's effect is on CPU temperature and stability.
 
Last edited:
Solution