Question 'Undervolt' Ryzen 3900X to reduce summer temperatures

[Fardringle]

I recently put together a Ryzen 9 3900X system with a Gigabyte Aorus X570 Elite Wifi motherboard, 32 GB of G.Skill RipJaws 3600Mhz CL16 RAM, and a Scythe Fuma 2 cooler. (plus my previous SSD and an EVGA GTX 1060 3GB)

Overall I'm very happy with the performance results. However, it's starting to get warmer here, and the already somewhat high temperatures in the system are just going to go higher as summer progresses, so I'm looking for the most efficient way to reduce heat without losing too much performance.

I already undervolted the GTX 1060 using this guide and dropped full load temps on the GPU by about 8C, which is a great start as it not only lowered the temps but actually boosted the GPU benchmarks a bit as well since the card was throttling itself a bit and not boosting all the way up to 1911MHz before the changes...

I'd really like to do something similar with the CPU if possible. Don't mind if I lose a small amount of performance if the temperature drop is good enough. Boosting performance and lowering temperature is probably too much to ask for on the CPU. I've found quite a few guides online for undervolting the CPU, but there are so many different options, and some of them seem to be excessively complicated and/or have pretty questionable results, so I'm hoping to get some good advice here.

It seems like the easiest option would be to simply lower the PPT to something below the stock 142W maximum, and with some trial and error I can probably find a suitable setting. But if there are better options, I'm definitely open to suggestions.

 

[UsandThem]

Maybe, but if I'm going to throw more money at it, I might as well return the Fuma and get something else...

Sure would like to see the temps they reported in that review, though!

Yeah, I think their ambient temps are just a TEENY bit lower than what you said you have.

 

[Fardringle]

Yeah, I think their ambient temps are just a TEENY bit lower than what you said you have.

Yup, and not the same CPU. But I can dream.

 

[DrMrLordX]

@Fardringle

Let me tempt you with some stuff I just now did with my 3900x and Prime95 SmallFFTs as a proxy for a DC workload (yes, some DC workloads can actually be worse).

I ran Prime95 SmallFFTs, and it instantly hit 3.9 GHz using 90% of PPT and ~62a. Ignoring the weird PPT behavior (probably a result of my undervolting, but who knows?), I set 3.9 GHz as my first target for a static OC. Temps during "default" operation hit 56.6C. Can I do better than that hand-tuning? It turns out, I can.

I opened up Ryzen Master, set a very conservative static OC to all cores of 3775 MHz @ 1.3v (1.14v actual according to CPU-z) and started tuning. First I bumped clocks up to 3900 MHz once I confirmed they would run that voltage. Then I started dropping volts using Ryzen Master with Prime95 running in the background.

Thus far I've gotten actual volts (again, according to CPU-z) down to 1.056v sustained @ 3900 MHz. Ryzen Master is reporting Tctl of ~53.4C, and all my workers are still running.

edit: lowest I can get without immediately killing a worker is 1.044v. Tctl is ~53.2C.

edit edit: looks like long-term stability is 1.056v where I had it above. Temps are slightly higher now @ 53.8C but holding steady. According to HWiNFO64, package power has dropped to 120W. I also tried running CBR20 with those settings and with higher voltage to check for false clocks, and it turned in a score of 7040 both times (3.9 GHz is not that fast folks), so I wasn't getting false clocks using the voltage settings that got me 1.056v in Prime95 SmallFFTs.

Anyway, bottom line is that you can beat the boost algo at a specific clockspeed. I dropped temps in this instance by about 3C, which ain't bad.

 

[Magic Carpet]

@DrMrLordX

Have you tried disabling individual threads/cores and how that affect power consumption?

 

[DrMrLordX]

@DrMrLordX

Have you tried disabling individual threads/cores and how that affect power consumption?

Nope. But I could if you'd like. Not sure if it's applicable to the OP since he's running DC apps that need all the thread capacity they can get.

 

[Magic Carpet]

Nope. But I could if you'd like. Not sure if it's applicable to the OP since he's running DC apps that need all the thread capacity they can get.

Yeah, maybe for another thread. It would be interesting to see how 3900x scales down power wise

 

[DrMrLordX]

Yeah, maybe for another thread. It would be interesting to see how 3900x scales down power wise

I can tell you that with my current RAM OC, uncore current consumption tops out at around 12a in a fully-loaded application. Core current has gone as high as 62-63a in stuff like Prime95. Assuming uncore current stays the same regardless of how many cores are fully-engaged, that should give you some idea of what happens when entire cores get shut down. But I'm not sure uncore/SoC current would remain static.

 

[moinmoin]

Isn't the point though to reduce heat production?

Better cooling also reduces heat production, since with the starting temperature being lower the CPU can run more efficient and stable.

 

[MrTeal]

Setting a temp limit achieves the same goal with less screwing around. It accomplishes the exact same thing only using temperature limit rather than power limit to induce throttling. Since the goal here is lower temps why not use temp limit rather than power? He can set the temp limit to whatever he wants.


Using a slight negative offset voltage with low LLC as @coercitiv suggested earlier should also work with the temp limit. It's a bit weird to dial in though. With clocks set on auto the CPU will clock stretch rather than crash when power starved. The clocks will appear normal or improved even though performance is reduced. You must test performance when under-volting with clocks set on auto. Typically any more than a -50-100mv offset will reduce performance. This is also extremely hard to properly test since the behavior can vary depending on workload. Since your only running a single workload this shouldn't be an issue. If clocks are set on auto you can't trust the clock speed the monitoring software is showing you when under-volting.

That's the long explanation for

What's your goal with trying to lower temperatures? Specifically, are you trying to lower CPU temperatures, or trying to reduce the amount of heat that's released into your room? They aren't the same thing.

If it's the latter, reseating the cooler or other improvements to cooling to reduce your temperatures won't help and you'll just need to lower the package power.

 

[mopardude87]

I never top over 80cel as of lately on this 3900x with a Dark Rock Pro4, that is with me running 97% BOINC, 5 tabs of chrome open and Folding@Home crunching on a 1080ti. I set the TDDP to 105 and LLC is at normal. This is a very typical load btw, maybe except one or two less tabs.

My room gets very toasty, my ac set at 76f but the room can hit upwards of 90. The crazy ass roomie got the boot but not before stealing my heat reflecting blanket so now that i moved back in here and without a curtain/blanket the sun beats right into my room. Oh karma got her cause the place she went to kicked her right the heck out. She tried to come back here.

Well i could upgrade to these then REALLY improve some things perhaps? I take function over looks any day.

Industrial Noise Control Acoustic Wall Blanket

Acoustic Industrial Noise Control - Sound Shield Noise Control Wall Blankets help reduce harmful noise from industrial machineries, compressors, and more.

www.strongarmstore.com

 

[Fardringle]

Well, that was a result I did not expect.

Out of curiosity, I set PPT to 105 Watts instead of 142, and that dropped average CPU temps to 72C (with a few spikes near 80C), which is certainly a more comfortable temperature range. Still higher than I'd like, but much better than 85C+ that I was getting at stock settings. All core average clocks are around 3.9Ghz with single core boosts up to around 4.4-4.5Ghz, compared to around 4.2Ghz average at stock and around 4.05-4.1Ghz with temps capped at 85C.

The surprising thing is that even with the lower average clock speeds, Passmark CPU scores with 105W PPT are only 3% lower than with the stock 142W PPT setting, and Cinebench R15 and R20 (all core) scores are only only 5-6% lower than the stock settings, and actually higher than I got when I tried setting PPT to 130W.

I expected to see temperature drop by lowering the power limit that much, but didn't expect to get higher scores than either the 130W PPT setting or the 142W/85C max temp setting...

It does hurt a little to lose about 5% of performance compared to stock settings, but I think I can live with that to get 10C (or more) lower temperatures. At least until the weather cools down in a few months so I can boost it back up to ludicrous speed again.

 

[aigomorla]

Oh i see you do BOINC, so i am guessing you will see a net positive in reducing vcore.
And yes it will be add up quite a bit that i believe you probably wont go back once you see the net savings in both electricity and heat.
Even the tiniest reduction compounded by 24 hours a day, leads to a massive reduction overall compounded 30 days a month, and 12 months a year.

 

[Fardringle]

Oh i see you do BOINC, so i am guessing you will see a net positive in reducing vcore.
And yes it will be add up quite a bit that i believe you probably wont go back once you see the net savings in both electricity and heat.
Even the tiniest reduction compounded by 24 hours a day, leads to a massive reduction overall compounded 30 days a month, and 12 months a year.

I wasn't even thinking about the savings on the power bill. I just wanted to get the temperatures down a bit. But yes, reducing CPU power usage by almost 40 watts on a system running at 100% 24/7 will definitely help keep the power bill under control.

 

[coercitiv]

It does hurt a little to lose about 5% of performance compared to stock settings, but I think I can live with that to get 10C (or more) lower temperatures.

Have you undervolted yet? You can probably get that performance back, either through a small undervolt, a more relaxed LLC setting, or a combination of both. I would start with looking at LLC, since that leaves light load and idle voltage intact, meaning it'll be much easier to test for performance & reliability.

 

[Fardringle]

Have you undervolted yet? You can probably get that performance back, either through a small undervolt, a more relaxed LLC setting, or a combination of both. I would start with looking at LLC, since that leaves light load and idle voltage intact, meaning it'll be much easier to test for performance & reliability.

I haven't tried undervolting yet, mostly because I can't find a voltage offset option anywhere in the Gigabyte BIOS, and don't see any helpful info when I try to search for it online. I'll probably feel dumb as soon as someone tells me where it is, but I'm not seeing it yet..

 

[mopardude87]

Oh i see you do BOINC, so i am guessing you will see a net positive in reducing vcore.
And yes it will be add up quite a bit that i believe you probably wont go back once you see the net savings in both electricity and heat.
Even the tiniest reduction compounded by 24 hours a day, leads to a massive reduction overall compounded 30 days a month, and 12 months a year.

I pay a set amount of rent per month, i do it out of consideration for my friend who rents me my room. That and well per clock 105 TDDP has been the best balance vs power usage vs clocks i seen yet. I love to balance things and this chip i think right now is perfecto! My room would only heat up further and for what? This thing does perfectly fine for what it is.

Edit: added picture just as proof for fun

https://imgur.com/a/hSLf7f2

 

[IEC]

ECO mode or use PBO to manually set PPT.

My 3950X on Eco mode loses nothing on lightly-threaded workloads and still has plenty of power for many-threaded workloads.

You can also hit very low idle usage by making sure to install the latest AMD chipset drivers, selecting the Ryzen Balanced power plan, and minimizing the "bloat" you run on Windows, especially monitoring utilities.

Doing nothing at the desktop: