• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

Noctua NH-U14S Getting 90C at 4GHz

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

tomdean

Member
Joined
Nov 26, 2011
I think my new system is too hot. I am getting 90C at 4GHz.
Am I trapping hot air in the case? Lots of exhaust at all the vent holes.

Gigabyte TRX40 AORUS Xtreme
Noctua NH-U14S TR4-SP3 with 2 Fans
Be Quiet! Dark Base 900 Full Tower
4 140mm fans (supply) front
1 140mm fan (exhaust) rear top
case has lots of vent holes on all sides
AMD 3970X Temperature, Degrees C

(I tried to paste a table, but, cannot seem to get a fixed font.) Attached a file.
Running Prime95, looking at temperature
Freq Tctl
4.03GHz 87.25 C
3.64GHz 76.25 C
3.63GHz 78.875 C
4.15GHz 60.0 C
4.14GHz 75.875 C
4.02GHz 90.5 C
4.01GHz 91.0 C
4.03GHz 90.5 C
4.02GHz 92.25 C
3.99GHz 94.75 C
3.95GHz 94.75 C
3.60GHz 77.25 C
3.62GHz 80.0 C
4.02GHz 89.5 C
4.02GHz 89.25 C
4.01GHz 90.125 C
3.97GHz 94.75 C
3.94GHz 94.75 C
3.59GHz 78.25 C <=== stopped prime95View attachment NH-U14S.txt
3.06GHz 45.625 C
3.38GHz 33.875 C

- - - Auto-Merged Double Post - - -

2nd try at attach:
View attachment 214133
 
OP
T

tomdean

Member
Joined
Nov 26, 2011
I went with the Noctua NH-U14S TR4-SP3 because it was stated to be as good as water cooling.
 

EarthDog

Gulper Nozzle Co-Owner
Joined
Dec 15, 2008
Location
Buckeyes!
I went with the Noctua NH-U14S TR4-SP3 because it was stated to be as good as water cooling.
That doesn't mean much. It's a high-end air cooler, indeed...but you're using a 280W CPU, dude. I'm not surprised in the least.

You can add another exhaust fan or two... that may help a couple of degrees(?), but otherwise, yes, this seems about right for the CPU sitting under it and the stress test you're running. Also, what specific test in P95 are you running? Small FFT? Blend? Large? That matters as the temps vary dramatically.
 
OP
T

tomdean

Member
Joined
Nov 26, 2011
That doesn't mean much. It's a high-end air cooler, indeed...but you're using a 280W CPU, dude. I'm not surprised in the least.

You can add another exhaust fan or two... that may help a couple of degrees(?), but otherwise, yes, this seems about right for the CPU sitting under it and the stress test you're running. Also, what specific test in P95 are you running? Small FFT? Blend? Large? That matters as the temps vary dramatically.

I do not remember what test I ran. It loaded all cores to 100%. results.txt has
[Tue Jul 13 14:14:08 2021]
Self-test 384K passed!
Self-test 384K passed!
Self-test 384K passed!
Self-test 384K passed!
Self-test 384K passed!
Self-test 384K passed!
Self-test 384K passed!
Self-test 384K passed!
Self-test 384K passed!
Self-test 384K passed!
...
 

EarthDog

Gulper Nozzle Co-Owner
Joined
Dec 15, 2008
Location
Buckeyes!
That data is not terribly helpful/telling to me, sorry. All P95 tests load all cores to 100%. Maybe walk through your process and figure it out as it's certainly good to know.

Typically if you're stress testing your CPU, you want to run Small FFT. If you want to test memory as well, run Blend. The latter (Blend) tends to run cooler than Small/Smallest FFT.
 
OP
T

tomdean

Member
Joined
Nov 26, 2011
I ran 'mprimes -m' and selected #16, torture test, 64 threads, small fft.
 

EarthDog

Gulper Nozzle Co-Owner
Joined
Dec 15, 2008
Location
Buckeyes!
Not sure what that is, bud. You should select Options->Torture Test and you select the # of threads and mode.

p95.jpg

That said, sounds like Small FFT, so, it's one of the hottest tests. 90C + that air cooler + P95 Small FFT = nobody being surprised. :thup:
 
OP
T

tomdean

Member
Joined
Nov 26, 2011
I tried to repeat the earlier measurements with altered cooling. I measured air flow and temps at intake and exhaust. I had pressure in the case.
I removed the case side panel.
The idle temp dropped 2 C.
I started prime95, small fft.
The temps at 3.5GHz were 9 C cooler. But, that was it.
The freq did not go above 3.6 GHz. The temp was 71 C. Ran for an hour like that.
Strange. This is Ubuntu 20.04. As far as I can tell the settings are the same.
Any ideas as to what is limiting the frequency increase?
 
OP
T

tomdean

Member
Joined
Nov 26, 2011
I found the problem, but, do not understand it.
I set the freq governor to performance and had freqs > 4GHz. As soon as I started prime95, the freqs dropped to 3.5GHz and stayed there. Same for governor ondemand.

I created an 'exe' that calculated sin,cos, tan in a loop, forever, using volatile variables. Looking at objdump -d, I can verify the loop.
I ran 61 copies of the 'exe' and verified that I had them on different CPUs
The freq was 4.15GHz to 4.20GHz and the temps were 54C to 56C.

The only change I made to hardware was to remove the case side panel.
My problem was not enough through air flow. I will add a couple 140mm exhaust fans at the top of the case.
The NH-U14S will handle more. I need to do BIOS OC to get higher freqs.

I attached the results. View attachment NH-U14S-open.txt
 

EarthDog

Gulper Nozzle Co-Owner
Joined
Dec 15, 2008
Location
Buckeyes!
Sooooo, you've limited the frequency and then changed stress tests (among other things) and temps are lower? In order to see if anything is fixed, shouldn't you run p95 without limiting the cpu? Do you have a baseline for the new test? I'm really confused at what you're trying to show and the methodology... maybe it's me? :chair:

Whatever "exe" you made clearly isn't stressing out the cpu like p95 does, though.
 
OP
T

tomdean

Member
Joined
Nov 26, 2011
No, I did not limit the frequency. For some unknown reason, when I run p95, the frequency is limited by the OS, not me. Before I started P95, the freq was 4.2GHz and as soon as I started P95, the freq dropped to 3.5GHz. I did not make any changes to the OS, just rebooted. Has to be something in runtime config. I can not find it.

The 'exe' I created uses 100% of whatever CPU it is on. Nothing else. Running 61 copies of the 'exe' gets the CPU to near 100% load (61/64). I ran with nice -n -20 to avoid context switches. P95 can not stress the CPU more than 100%. The biggest difference is the 'exe' I made does not stress loading cache. It easily fits in cache and does not trigger a miss and cause a fetch. I can not get CPU freq above 4.2 GHz with this OS configuration, without doing BIOS OC. That is next.

What I wanted to test was the air flow. I ran the test for more than an hour. And, removing the side panel shows that I have some restriction that is limiting flow. More fans.

And, the Noctua NH-U14S TR4-SP3 will handle the 3970X to more than 4.2GHZ at 100% load and keep the temps below 60C.
 

EarthDog

Gulper Nozzle Co-Owner
Joined
Dec 15, 2008
Location
Buckeyes!
The 'exe' I created uses 100% of whatever CPU it is on. Nothing else. Running 61 copies of the 'exe' gets the CPU to near 100% load (61/64). I ran with nice -n -20 to avoid context switches. P95 can not stress the CPU more than 100%. The biggest difference is the 'exe' I made does not stress loading cache. It easily fits in cache and does not trigger a miss and cause a fetch. I can not get CPU freq above 4.2 GHz with this OS configuration, without doing BIOS OC. That is next.
It's pretty clear whatever you're doing isn't close to the same stress level as p95. Just because it says 100% doesn't mean its doing the same thing tickling the same parts. Take a look at power use when running your exe and p95. ;)

And, the Noctua NH-U14S TR4-SP3 will handle the 3970X to more than 4.2GHZ at 100% load and keep the temps below 60C.
Maybe with your exe... but not with p95.
 
OP
T

tomdean

Member
Joined
Nov 26, 2011
From p96: "Prime95 has been a popular choice for stress / torture testing a CPU since its introduction, especially with overclockers and system builders. Since the software makes heavy use of the processor's integer and floating point instructions, it feeds the processor a consistent and verifiable workload to test the stability of the CPU and the L1/L2/L3 processor cache. Additionally, it uses all of the cores of a multi-CPU / multi-core system to ensure a high-load stress test environment."

The application I wrote does all this and since it does not do anything other than long double floating point and long integer instructions, it may stress the CPU more than p95. There are no I/O actions to give the CPU a break.
If you look at the CPU usage of p95, it varies and is not always 100%. This is when p95 is writing to files like results.txt.
My application is ALWAYS at 100% since it never does any I/O.
 

freeagent

Member
Joined
Sep 15, 2004
Location
Winnipeg!
Your CPU speed is not limited by the OS when you run P95, it is limited by the socket.. or else your CPU will pull enough power to commit suicide? I don't recall the reason for limiting socket power. AM4 users face the same type of power limits.

If your CPU is running cooler and faster with your exe its because your CPU isn't running as hard as it was with P95, regardless of what you think your code is doing. Also, the cooler you can keep it, the faster it will run. The fact that it is dropping all the way down to 3500 tells me it is not a happy camper.
 

EarthDog

Gulper Nozzle Co-Owner
Joined
Dec 15, 2008
Location
Buckeyes!
If it's the same load, why are the temperatures over 30C less? How is it possible your exe is doing more yet running so much cooler? The data doesn't seem to match what you're saying.
 
OP
T

tomdean

Member
Joined
Nov 26, 2011
Ah. That is what is missing. I changed the cooling.

Measurements with an air flow meter indicated there was positive pressure in the case. Hot air was trapped in the case. And, most likely, air flow through the Noctua NH-U14S TR4-SP3 was reduced.

I removed the case side panel.

Air flow through the front panel fans significantly increased, sort of proving that before removing the side panel, there was pressure in the case.

So, I need to do something to improve air flow with the case closed. I do not want a negative pressure in the case. The Noctua NH-U14S TR4-SP3 is a poor design in that the hot exhaust air from it is not vented outside the case. To be cooler, get rid of the hot air...
 

mackerel

Member
Joined
Mar 7, 2008
Only just saw this thread. So another quick description of Prime95:

To describe the bulk of the execution workload Prime95 does on modern CPUs, consider it FP64 (double precision float) heavy. If you have good AVX2 support (Haswell or newer, Zen 2 or newer), or even better AVX-512 (Skylake-X HEDT or newer, Ice Lake mobile or newer, Rocket Lake desktop or newer), it will chew through those at an insane rate. Where the FFT size comes in relates to the data size in use by each thread. Rule of thumb is FFT size multiplied by 8 is the data size. e.g. 128k FFT = 1MB data size. This isn't perfect as there is also some non-FFT data, but seems to work well enough in practice. If the effective cache size (L3 only for inclusive cache CPUs, total L2+L3 for others) of the CPU is big enough to hold all the data by all threads then the CPU cores will get a heavy workout. If the cache is exceeded, then memory bandwidth comes into play. Memory latency is far less important since it does a good job at prefetching. As far as I'm aware, Prime95 is possibly the heaviest CPU real world workload, with only Linpack variations coming close.

Specifically with AMD CPUs, especially since Zen 2 when AMD finally mad a half decent FPU, at stock they run to a fixed power limit. As such, everything I just wrote above largely doesn't matter because you're running at the power limit a lot of the time when running all core. 280W of Prime95 is not hotter than 280W of Cinebench for example. Prime95 is still a heavier workload, which is why you might see lower clocks running that than lower intensity loads like Cinebench.

Since Zen 2 I think it largely pointless running fixed clock overclocks on AMD CPUs. At best you can gain in some workloads you care about by sacrificing stability in workloads you might never run like Prime95. Power limit increasing is about the only smart way to do it, but the gains are minimal as you're running in a very inefficient part of the curve before you even start.
 

EarthDog

Gulper Nozzle Co-Owner
Joined
Dec 15, 2008
Location
Buckeyes!
Good p95 info! In short, do you agree that temps are normal for the p95 load and his cooler, maMac?


Tomdean, why do you not want more exhaust than intake? The dust difference is negligible in most cases. But that said, if your temls dropped about 2C removing the side panel, it doesn't seem like it's struggling for airflow. I'd still add at least one fan for exhaust, however. Cases aren't really sealed so 'pressure' differences are negligible (air flow, CFM, isn't ;)).
 

mackerel

Member
Joined
Mar 7, 2008
Good p95 info! In short, do you agree that temps are normal for the p95 load and his cooler, maMac?

I don't have experience of the CPU so hard to say, but assuming 280W, then I'd certainly expect it to get very hot with ANY load hitting the power limit. P95 or not doesn't matter. Zen 2 was and likely remains a high power density CPU. In theory at 280W for the CPU you're looking at 70W per CCD (ignoring the IOD for now). With single CCD Zen 2 they were 88W PPT, and two CCD Zen 2 was 142W total (71 each?), again ignoring IOD so this isn't exact. Power density on the CCDs is similar ball park. Separate question then how effective the U14S is at shifting that heat. Personally I'd prefer something beefier but I've not looked at TR4 coolers to see what's available.