I think SpeedFan and Fan Control can adjust fan speeds based on various sensors but haven’t used either and doubt either work for Linux.
Edit: overlooked you asked about hardware solution. I don’t know of anything.
1 Like
I was wondering about air cooling as well, so I ran some tests of a NH-D15 vs a Corsair H150i RGB Pro XT with the CINEBENCH R23 10-minute throttling test. Surprisingly, the D15 performed as well or slightly better than the H150i. The simple summary is that the D15 is fine for cooling a 7950X.
The H150i was configured as an intake with 3 NH-A12x25 fans and tested at 50% fan, 100% fan, and 100% fan with max pump speed. The D15 was configured with the included NF-A15 and a NH-A12x25 front fan (for RAM/case clearance) and tested at 50% fan and 100% fan.
Results:
H150i 50% fan - score 36,931
H150i 100% fan - score 37,285
H150i 100% fan max pump - score 37,351
D15 50% fan - score 36,853
D15 100% fan - score 37,732
Steady state temperatures were all at the 95C boost target. Average steady state clocks ranged from around 4915 MHZ (D15 50% fan) to around 5070 MHZ. Power draw was from 194W (D15 50% fan) to 215W.
One performance advantage the H150i has is that the initial boost clocks and power draw last several seconds longer, probably due to the extra thermal capacity of the water. It also performed a bit better at 50% fan, which makes sense since it has additional radiator area.
I’m not 100% sure why the D15 score slightly better - it could just be down to slightly better thermal transfer, run to run variance, or maybe the iCue service sucking up CPU time.
I’d be curious to see what you get with a safe undervolt. IIRC someone mentioned -10 all core was a decent value for starting? Though DYOR.
I’m running curve optimizer right now to see what it recommends for undervolting and maybe tweak from there. Dropping PPT to 175W results in the temp dropping well below 95C with only a relatively small drop in performance.
The issue with cooling seems to be getting heat off the die(s) rather than total thermal dissipation. Even a NH-U14 can dissipate 280W from a Threadripper and Noctua rates it lower than the NH-D15.
A gentle reminder.
If your daily workloads are not hitting 143W, 175W or 230W, perhaps you don’t have to be overly worried about optimizing for the ‘peak’ operational point. Because you’ll hardly hit it, use it.
If your daily workloads are hitting those high-power operational points, Cinebench R23 might not be a good benchmark to help you test and verify your optmization. Because your workloads could be more computationally intensive than Cinebench R23.
Just my 0.02 from wasting tonnes of time dealing with AMD’s precision boost algorithm in the past couple of years.
2 Likes
Yes iCue sucks performance for sure.
The Ryzen Master auto curve optimizer recommended -30 for every core, after running for about 2 hours, so I went with that.
Cinebench R23 10 minutes test improved to 38,932 (and 39,534 for a single render), but temps still capped at 95C since the CPU just boosted higher - 5,225 MHZ all-core average.
The best way to cap temps is probably to just set the temperature target or reduce PPT.
I’m not sure the -30 is 100% stable yet, but so far it has passed every stress test I have tried.
1 Like
Use CoreCycler or something similar to load test each core one at a time.
Yep, the Ryzen Master auto curve optimizer in per-core mode looks like it cycles through each core, but more tests are always better.
I usually just run stock for maximum stability, but just having fun playing around with what the hardware is capable of.
2 Likes
You’ll have a golden sample if it is, CPU profile benchmark on 3D mark demo is a quick and dirty way of testing instability (not stability)
2 Likes
Corecycler, y-cruncher all tests, occt stability certificate, blender benchmark
2 Likes
As I comprehend it AMD just decided that 95C was a good temperature to run their CPU. If you want it to run cooler you can tune it to do so with no noticeable loss of performance. Plenty of videos explaining how to do this. However there is no need to bother. An AIO is very easy to install and is quite effective, I’d recommend one for the higher power CPUs but not essential.
1 Like
-30 all core isn’t quite stable since it fails the y-cruncher stress test after a while even at JEDEC memory timings. Still, stress testing was useful because it exposed some failures with the memory timings I had set. I want to pick up some DDR5 Unbuffered ECC DIMMS with Hynix memory chips.
1 Like
On Linux I have not had luck undervolting. When I was running my 3900X under windows I could get the core V down to 1.4 but my Linux workstation was 1.45V min period.
This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.