[SOLVED] Temperature issue with stock AMD Ryzen 5950X ?

[Wartem]

My question is why my stock CPU gets hot, only under low load.
The temperature quickly changes. The GPU is idling. There's good ventilation.
I no software installed, changing something like Dragon Center.
Windows 10.

Loads/Utilization (temperatures from CPU Package):
Idle: 38 degrees C.
1-5%: 40-50C.
10%: 60C - Unreal Editor.
15%: 75C - (Quickly/instantly). Noisy. Using Boinc or GTAV.
20%: 86C - (Quickly/instantly). Noisy. Using Boinc.
30%: 77C - (Quickly/instantly). Noisy. Using Boinc.
40%: 70C
50%: 68C
60%: 67C
70%: 65C
90%: 62C
100%: 58-63C Max load, all cores (Using Boinc and Cinebench). Goes down in temperature from above %, in just a few seconds. Silent fans.
100%: 62-68C using Cinebench single core.


Shouldn't the temperature be lower under lower load?


CPU: 5950x (Stock)
Cooler: Noctua NH-U12A
64 GB RAM.
Big case, good airflow. It doesn't get time to get hot from for example the GPU (idling) anyway since the temperature changes on the CPU package are close to instant.

Thanks in advance.

 

[drea.drechsler]

Get HWInfo64 to monitor temps. Look for a CPU Die (average) in the sensors and watch that; there's probably two...one for CCD1, one for CCD2. In particular, look at the AVERAGE column as that's the average across time of the average of the sensors across the die.

Ryzen's very different from prior CPU's in that it reports instantaneous hotspot sensor temp from one of dozens scattered across each die. Since it also boosts eagerly from any light load (simple as a mouse movement) there's always a hotspot lighting off to register a temp spike and...you see it instantly. The CPU die(average) is averaging the hot spots across the die to help with that. But that's instantaneous and thermal output is really a product of work performed across time and the AVERAGE column shows the average of the averages across time which is the best indicator of thermal output from the CPU die.

 

[CountMike]

My question is why my stock CPU gets hot, only under low load.
The temperature quickly changes. The GPU is idling. There's good ventilation.
I no software installed, changing something like Dragon Center.
Windows 10.

Loads/Utilization (temperatures from CPU Package):
Idle: 38 degrees C.
1-5%: 40-50C.
10%: 60C - Unreal Editor.
15%: 75C - (Quickly/instantly). Noisy. Using Boinc or GTAV.
20%: 86C - (Quickly/instantly). Noisy. Using Boinc.
30%: 77C - (Quickly/instantly). Noisy. Using Boinc.
40%: 70C
50%: 68C
60%: 67C
70%: 65C
90%: 62C
100%: 58-63C Max load, all cores (Using Boinc and Cinebench). Goes down in temperature from above %, in just a few seconds. Silent fans.
100%: 62-68C using Cinebench single core.


Shouldn't the temperature be lower under lower load?


CPU: 5950x (Stock)
Cooler: Noctua NH-U12A
64 GB RAM.
Big case, good airflow. It doesn't get time to get hot from for example the GPU (idling) anyway since the temperature changes on the CPU package are close to instant.

Thanks in advance.

That's inadequate cooler for a 16 core beast, just not enough mass to keep temps steady, on top of that, Ryzen 5000 series are hotter and allowed higher temps without throttling. As i is, you are getting temps better than expected for such cooler. Ryzen and most new CPUs are made to react to load fast and temps raise and fall at same cadence.
Looking at those temps, I suspect you are not getting best performance out of it, PBO or boost is probably not enabled, or it would get much hotter. Similar cooler couldn't keep in check even my 3700x. You need a 360 liquid cooler to get full value out of 5950.

 

[Wartem]

That's inadequate cooler for a 16 core beast, just not enough mass to keep temps steady, on top of that, Ryzen 5000 series are hotter and allowed higher temps without throttling. As i is, you are getting temps better than expected for such cooler. Ryzen and most new CPUs are made to react to load fast and temps raise and fall at same cadence.
Looking at those temps, I suspect you are not getting best performance out of it, PBO or boost is probably not enabled, or it would get much hotter. Similar cooler couldn't keep in check even my 3700x. You need a 360 liquid cooler to get full value out of 5950.

I've had my CPU on 100% all cores all day with silent fans.

I just read that the issue or part of it is that the CPU is allowed to boost higher on fewer cores and it needs the higher voltages to maintain stability. That means heat.

CPU:
20% utilization on all cores (it changes to these values quickly (seconds) and stay there until I change the load):
85-87C (CPU package).
125-142W
About half of the cores at 4700Mhz, the other half at 3700, shifting rapidly.
~1.37V

100% utilization on all cores (running like this for maybe 10 hours):
~63C (CPU package).
~130W,
~3700Mhz,
~1.0V

__

@drea.drechsler
Regarding CPU Die (average): I will test this later and let it run for a while. Thanks for the info!

Edit:
Short test with temp at 20% CPU load:
CPU Package: 85-89C
CPU CCD Max: 80-92C
CPU Average: 60-75C
CPU CCD1: 73-92C
CPU CCD2: 43-68C

Difference in fans: Only difference from CPU 100% to CPU 20% was apparently one fan which changed in speed from 70 to 100% speed. It sounded like more. The rest of the fans are at 60% in both scenarios.
I guess I've never heard what it sounds like with all fans at 100%.

Edit2:

Decibel levels at 100% CPU utilization:
Placing my phone on the chassi, in the middle on top: 29-32 decibel. Same when placed 1 meter from the PC, where I sit.

The output air isn't especially hot. The chassi is cool on touch.
I've got another PC with same chassi and CPU cooler but with 5600x and 6800 xt.. and that graphics card (compared to the CPU) gets really hot and loud under pressure! Perfect to use during cold winters.

 

[drea.drechsler]

....
Short test with temp at 20% CPU load:
CPU Package: 85-89C
CPU CCD Max: 80-92C
CPU Average: 60-75C
CPU CCD1: 73-92C
CPU CCD2: 43-68C

Difference in fans: Only difference from CPU 100% to CPU 20% was apparently one fan which changed in speed from 70 to 100% ...

Try a proper heavy CPU load...CinebenchR20 or R23, multithread. If that CPU Average is holding around 75C with that you're good as gold.

It's not uncommon for Ryzen 5000 CPU's to hit around 90C in use. AMD's told us it's by design to extract full processing potential of the CPU since that's the line in the sand where the boost algorithm will pull back. The silicon is designed to handle it, it's only the owners that aren't.

 

[Wartem]

Ok, I don't feel like 100% CPU load ever has been an issue. Like I said, I've been using it at 100% all day. The issue was the 20% load. Maybe I misunderstood you.
But I ran Cinebench R23 "CPU (Multi Core)", 24263 pts anyway:
(Did a reset of the monitor's max/min values before running):

Max temp from "CPU CCD Average" was 61,4C.

 

[drea.drechsler]

Ok, I don't feel like 100% CPU load ever has been an issue. Like I said, I've been using it at 100% all day. The issue was the 20% load. Maybe I misunderstood you.
But I ran Cinebench R23 "CPU (Multi Core)", 24263 pts anyway:
(Did a reset of the monitor's max/min values before running):

Max temp from "CPU CCD Average" was 61,4C.

20% load appears an issue because the CPU is freely boosting to max clocks still...and so you get instantaneous hot spot temp spikes that are high. But that's insignificant since actual thermal output (CPU Die(average) ) is still very low.

At a 100% heavy workload the CPU has stopped boosting freely, if at all, and so the instantaneous hotspot temperature spikes aren't occurring. But the cores are now working constantly, all of them, so actual thermal output can be extreme even if not spiking so high...and so that CPU Die(average) reading should rise higher. If it's holding at 61.4C through a CBR23 BM run then IMO you're very, very good for a 5950X.

All of this can make sense when you compare this to lighting a match in the room. Even with it's and extremely high temperature in the flame it's incapable of heating up the room until a thousand are lit all at once. That's the difference between temperature and thermal output.

 

[Wartem]

20% load appears an issue because the CPU is freely boosting to max clocks still...and so you get instantaneous hot spot temp spikes that are high. But that's insignificant since actual thermal output (CPU Die(average) ) is still very low.

At a 100% heavy workload the CPU has stopped boosting freely, if at all, and so the instantaneous hotspot temperature spikes aren't occurring. But the cores are now working constantly, all of them, so actual thermal output can be extreme even if not spiking so high...and so that CPU Die(average) reading should rise higher. If it's holding at 61.4C through a CBR23 BM run then IMO you're very, very good for a 5950X.

All of this can make sense when you compare this to lighting a match in the room. Even with it's and extremely high temperature in the flame it's incapable of heating up the room until a thousand are lit all at once. That's the difference between temperature and thermal output.

Thanks for explaining. I was afraid something was broken. I should just ignore the increased fan speed then at and around 20% CPU utilization. I have to run Boinc in the background all the time (when doing something which gives at least some load). And thereby making sure the CPU is less loud by making it 100% busy.

 

[drea.drechsler]

Thanks for explaining. I was afraid something was broken. I should just ignore the increased fan speed then at and around 20% CPU utilization. I have to run Boinc in the background all the time making sure the CPU is 100% busy.

I'd set up a custom fan curve with a base fan speed that's barely audible up to a fairly high temperature. The good thing about Noctua coolers is their fans: even at 100% they are rarely annoying so that should be fairly high speed. At something around 75 or 80C, experiment with it a little, start raising it towards max. Also with very long hysteresis times (time to rise/time to fall) as a steady, not pulsing, fan speed is much less annoying.

But this is also where a 360mm AIO is very good since you could set it's fans to only kick up when temps are 85-90C. That's acceptable since you're really using the enormous capacity of the liquid to absorb heat and only need the fans when it's getting thermally saturated.

 

[mamasan2000]

Try a proper heavy CPU load...CinebenchR20 or R23, multithread. If that CPU Average is holding around 75C with that you're good as gold.

It's not uncommon for Ryzen 5000 CPU's to hit around 90C in use. AMD's told us it's by design to extract full processing potential of the CPU since that's the line in the sand where the boost algorithm will pull back. The silicon is designed to handle it, it's only the owners that aren't.

Cinebench isn't a stresstest. It's a benchmark.
You should run Prime95.
With PBO enabled, the CPU acts like an ADHD patient. I'm not into that.
So I run a static overclock, same clocks as PBO in multicore loads but way less voltage and therefor lower temps, at every load.
When am I ever going to see a singlecore load in a game? Maybe the Loading screen but other than that? There's always something running on another thread, sound, loading light assets, whatever it is. Most games I play use 6+ threads.

 

[drea.drechsler]

Cinebench isn't a stresstest. It's a benchmark.
You should run Prime95.
With PBO enabled, the CPU acts like an ADHD patient. I'm not into that.
So I run a static overclock, same clocks as PBO in multicore loads but way less voltage and therefor lower temps.
When am I ever going to see a singlecore load in a game? Maybe the Loading screen but other than that? There's always something running on another thread, sound, loading light assets, whatever it is. Most games I play use 6+ threads.

Prime95 is heinously excessive, a true power virus that's utterly unrepresentative of anything real world. If anyone MUST run it on modern hardware then they must also run it with AVX disabled. IMO it's only purpose is a curiosity and when used knowledgeably to fully de-rate a system to engineered standards for high-availability and/or extreme environment operation.

Some people like de-rating their machine for their daily runner, that's fine for them. Most just want stability and safe high performance for what they do on a daily basis. In addition to being a good benchmark of high core count CPU's, CBR20 and R23 is also a good stress tester since it very much represents the worst case with a realistic mix of AVX and other instructions in a real world processing load of rendering an image.

But even so, when Ryzen's left in stock conditions, both VCore and clocks on AUTO, it will run P95 all day long even if at high temp's since the boost algorithm is keeping it safe. It's with the static overclocks when the safety mechanisms of the boost algorithm can no longer function that things get dangerous and de-rating may be worthwhile.

Even heavily multithreaded games depend on one primary thread. When CPU performance is the limit to FPS that one primary thread is the limit to the CPU's performance. So keeping that thread fixed at a clock speed lower than what it could boost to when left to it's own devices is just limiting performance. But, luckily, most gaming is GPU bound anyway so you're not likely to ever see it.

 

[mamasan2000]

Prime95 is heinously excessive, a true power virus that's utterly unrepresentative of anything real world. If anyone MUST run it on modern hardware then they must also run it with AVX disabled. IMO it's only purpose is a curiosity and when used knowledgeably to fully de-rate a system to engineered standards for high-availability and/or extreme environment operation.

Some people like de-rating their machine for their daily runner, that's fine for them. Most just want stability and safe high performance for what they do on a daily basis. In addition to being a good benchmark of high core count CPU's, CBR20 and R23 is also a good stress tester since it very much represents the worst case with a realistic mix of AVX and other instructions in a real world processing load of rendering an image.

But even so, when Ryzen's left in stock conditions, both VCore and clocks on AUTO, it will run P95 all day long even if at high temp's since the boost algorithm is keeping it safe. It's with the static overclocks when the safety mechanisms of the boost algorithm can no longer function that things get dangerous and de-rating may be worthwhile.

Even heavily multithreaded games depend on one primary thread. When CPU performance is the limit to FPS that one primary thread is the limit to the CPU's performance. So keeping that thread fixed at a clock speed lower than what it could boost to when left to it's own devices is just limiting performance. But, luckily, most gaming is GPU bound anyway so you're not likely to ever see it.

What's 'de-rating'?
If Prime95 is too heavy for your blood, at least run OCCT. Cinebench is not a stresstest. It is also not a stability test. Either your system is stable under every load or it's not stable. Like for instance, converting a video in AviDemux, can take 30+ minutes. You want to be stable for that. Cinebench is like a 30 sec test that barely heats up anything and mainly fits in your L3 cache as I understand it. A workload you will probably never see in the real world.
It's like being stable in 3Dmark while all your games are crashing because of the GPU OC. Are you then stable?

 

[drea.drechsler]

What's 'de-rating'?
...

De-rating is when you tune down performance to meet certain sub-spec performance parameters...like a lowered operating temperature and full stability even with total fan failures. It means the system will operate at less than rated performance with lowered clocks and voltage, but just fast enough to do what's needed. That might be done when installing in a low-airflow chassis or in a confined space, like a cubby hole in a desk or equipment cabinet, where it accumulates a few inches of insulating dust before someone gets around to checking on it. P95 is good at over-extending the system to see if you've got it far enough. We used something similar for setting up systems we installed in factory floor equipment because our manager refused to spend the huge sums the equipment OEM wanted for their PC controllers. Which was simply an already de-rated system built around an "industrial" motherboard.

I'm perfectly content using cinebench and leaving system in auto. I don't like pushing to the edge of dynamic stability like you have to with a static overclock on Ryzen to see any improvement over what PBO provides, with the boosting algorithm left in place to protect the CPU should someone do something stupid...like run Prime 95, small FFT, without disabling AVX.

EDIT add: and my point really is: if you do leave it in AUTO then sure, temps will get hot running P95 but that proves nothing since the CPU is DESIGNED to run up to 90C whereupon the algorithm does it's thing and limits boosting. The only thing that matters is that should it be getting that hot is it still performing on-spec since you can't look at clocks (way to dynamic)to gain any idea of that. You need a proper benchmark, with a lot of history to know if your cooling is OK. CBr20 or r23 fits that bill too so if it's running with no higher than 75C AVERAGE temps in a STRESS test mode then it surely should be pulling down good BM numbers too.

 

[Wartem]

I'd set up a custom fan curve with a base fan speed that's barely audible up to a fairly high temperature. The good thing about Noctua coolers is their fans: even at 100% they are rarely annoying so that should be fairly high speed. At something around 75 or 80C, experiment with it a little, start raising it towards max. Also with very long hysteresis times (time to rise/time to fall) as a steady, not pulsing, fan speed is much less annoying.

But this is also where a 360mm AIO is very good since you could set it's fans to only kick up when temps are 85-90C. That's acceptable since you're really using the enormous capacity of the liquid to absorb heat and only need the fans when it's getting thermally saturated.

Great advice, thanks.

I replaced the default rear exhaust 3-pin fan with a 4-pin Noctua NF-A14,
connected to a SYS_FAN connector. I also made custom fan curves and the PC is more silent now. If the room is otherwise silent, I can hear the fans when the CPU is at 100% utilization. So it's not like running a GPU at 100%, to put it lightly.

I'm happy with the results but will learn more in the future.

Under 20% load I now get CPU Package temp at 86-90C.
"CPU CCD Average" gets up to max 76C.

New test with Cinebench R23 CPU (Multi Core), 24491 pts:
"CPU CCD Average" gets up to max 63,8C.

 

[MR_BROCK]

Thanks for explaining. I was afraid something was broken. I should just ignore the increased fan speed then at and around 20% CPU utilization. I have to run Boinc in the background all the time (when doing something which gives at least some load). And thereby making sure the CPU is less loud by making it 100% busy.

I have a 5950x and unless my fans are set agressively on the fan curve it get hotter than yours under full load im running a 3060 ti underneath it as well . I reach like 75c . But its normal from what i read. And it spikes the same.