[SOLVED] Undervolting 3800X

[andy0x2a]

I have a 3800X with PBO enabled (motherboard limits, 2x scalar). I am seeing 4600Mhz max boosts, and average boosts under load around 4200. ambient of 24, idle 29, cinebench R20 5025-5050 @75C

I have tried a -0.10V offset in the BIOS, which resulted in a drop in temperature under heavy load of 5C, and increased cinebench score by 25-50 points. Max boosts dropped down to 4475/4500. I like the drop in temperature, and it seems to have a small performance increase with PBO. I am running some system stability tests now (p95), and it appears stable with the small offset.

However, I have seen several articles/threads on various forums that warn against negative voltage offsets, so I was wondering if there was any advice on this forum around that.

 

[Karadjgne]

Ppl have been undervolting cpus for years. It's common practice because Every cpu is different and has different voltage/current demands for stability. So intel/amd sets stock voltages high enough to cover any possible contingency, which may be just over what your cpu needs or way over what it needs. Using offsets changes VID, not vcore. VID is what the cpu demands from the VRM's, vcore is what the cpu gets. SVI2 TFN is what the cpu is actually using regardless of what it's demands or supply is.

 

[drea.drechsler]

I have a 3800X with PBO enabled (motherboard limits, 2x scalar). I am seeing 4600Mhz max boosts, and average boosts under load around 4200. ambient of 24, idle 29, cinebench R20 5025-5050 @75C

I have tried a -0.10V offset in the BIOS, which resulted in a drop in temperature under heavy load of 5C, and increased cinebench score by 25-50 points. Max boosts dropped down to 4475/4500. I like the drop in temperature, and it seems to have a small performance increase with PBO. I am running some system stability tests now (p95), and it appears stable with the small offset.

However, I have seen several articles/threads on various forums that warn against negative voltage offsets, so I was wondering if there was any advice on this forum around that.

Small negative offsets have been helpful for tweaking it to 'the most'. I'm using a -0.0125 offset for my PBO'd 3700x, for instance. But the problem comes in with large undervolts, and absolutely avoid fixed voltages, even if to undervolt, unless you really know what you're doing.

The problem with undervolting too far is something called 'clock compression'. The CPU is still reporting high clocks but performance actually suffers especially with lightly threaded workloads. My suggestion is always include a CB20 single thread test to see if you're hurting that. Lightly threaded performance is most critical with gaming so you really don't want to hurt that.

 

[andy0x2a]

I'm using a -0.0125 offset

Thanks for the feedback. I have tried -0.050 and -0.10, I have found temperature multicore improvement with -0.10. However, it seems that my offset is 8x higher than yours, which concerns me.

But the problem comes in with large undervolts,

Is a -0.10V considered a large undervolt? I am going to test with one similar to yours at 0.0125, and see if that makes a difference.

 

[drea.drechsler]

Is a -0.10V considered a large undervolt? I am going to test with one similar to yours at 0.0125, and see if that makes a difference.

If I used a -0.100 offset I'm confident ST performance would tank with my system...it did at only -0.0500 offset when I worked it up to -0.0125. From what I can tell, using a negative offset does help at all-core heavy loads by lowering temperature, as you described, but too much hurt at single thread loads just wasn't tolerable to me.

It's not going to be harmful, it just won't hit or sustain high boost clocks for light bursty loads. That's the exact kind of processing loads that many games present.

BTW: since you're using PBO, you might try these settings instead of just sliding them to the max: PPT = 333, TDC and EDC both = 230. For some reason that seems to work very nicely.

 

[andy0x2a]

but too much hurt at single thread loads just wasn't tolerable to me.

Thanks for your input.

I have run a quick test with Single Threaded cinebench 20. Here are my current results:

Voltage Offset	Cinebench ST
0/Stock	513
-0.0125	508
-0.0500	508
-0.1000	511

What is interesting here, is that I see a definite uptick at -0.10V, and that is strange. I also see a loss of 1.00% ST performance at -0.05, but a gain of 1.26% MT, as well as lower temperatures by 5C.

 

[drea.drechsler]

Thanks for your input.

I have run a quick test with Single Threaded cinebench 20. Here are my current results:

Voltage Offset	Cinebench ST
0/Stock	513
-0.0125	508
-0.0500	508
-0.1000	511

What is intersting here, is that I see a definite uptick -0.10V, and that is strange. I also see a loss of 1.00% ST performance, but a gain of 1.26% MT, as well as lower temperatures by 5C.

that is WEIRD...wow...I've never tested that much of an offset. maybe i will now! But, it's also true that motherboard VRM performance can play in this a lot, so your board is different and will doubtless react differently across loads. That's why you have to tweak to find what's optimum for your rig.

But also, since you're the tweaking sort you might like to try out those PBO settings and see if they do anything.

 

[andy0x2a]

But also, since you're the tweaking sort you might like to try out those PBO settings and see if they do anything.

I have done a quick test, and these are my results with PBO-4, no voltage offsets.

	R1	R2	R3	AVG	DEV
MB MAX	4999	4999	5018	5,005.33	10.97
CUSTOM	4999	5006	5016	5,007.00	8.54

This appears to have no significant difference from motherboard limits.

I am going to try with different offsets and your recommended settings, as well as various other offsets/settings combinations.

 

[drea.drechsler]

I have done a quick test, and these are my results with PBO-4, no voltage offsets.

	R1	R2	R3	AVG	DEV
MB MAX	4999	4999	5018	5,005.33	10.97
CUSTOM	4999	5006	5016	5,007.00	8.54

This appears to have no significant difference from motherboard limits.

I am going to try with different offsets and your recommended settings, as well as various other offsets/settings combinations.

So your's isn't following the same as others. Miy 3700x gave me a definite improvement in MT and ST scores - ST went from 4965ish to 5155ish and ST from 490 to 512. The MT score varies quite a bit; a cold boot in the morning and I can get 5193. But in the middle of the day, three runs with a minute rest in between will start 5165 then drop to 5145 by the third.

BTW..when I increased offset to -.1000 my MT score dropped 500 pts...from 5155ish to 4653. I didn't really bother with ST testing (takes so long) as it's immaterial at that point. So there is definitely a difference, probably down to VRM performance over load, in how our boards treat this.

 

[andy0x2a]

core dropped 500 pts...from 5155ish to 4653.

Uh.... wow, that is quite a drop! It just goes to show how different these setups can be.

cold boot in the morning and I can get 5193.

I am quite impressed at your score, I have never been able to get close to 5200. But I agree, I have seen a lot of variation depending on time-of-day/ambient temperature.

Based on your results, I am going to play with PPT/EDC/TDC a bit more, see if I can find a sweet spot.

 

[andy0x2a]

Also, I just ran -0.125V.... which gives a ST performance of 518. (and a MT score of 5062)

I am utterly confused by this.

 

[drea.drechsler]

...
Based on your results, I am going to play with PPT/EDC/TDC a bit more, see if I can find a sweet spot.

If you don't mind tweaking and really want to find the sweet spot for PBO with your rig then check out this thread:

https://www.overclock.net/forum/13-amd-general/1741052-edc-1-pbo-turbo-boost.html

The experienced Ryzen overclockers have been complaining how PBO boosting works with Ryzen 3000 compared to how it did with 2000 CPU's in that does not follow the PPT/TDC/EDC settings with any coherency. As it turns out, there is a bug in AGESA 1004b PBO. They've found a way to at least get better PBO performance, even if it's still not exactly 'coherent' to the settings.

And lastly: do not forget that CPU boosting is extremely temperature sensitive. CB20 scores will vary quite a bit through the day on my system, especially the MT scores.

 

[andy0x2a]

If you don't mind tweaking and really want to find the sweet spot for PBO with your rig then check out this thread:

Thanks, I will take a look. I quite enjoy tweaking my computer, it really is an enjoyable experience.

I definitely know what you mean about temperature sentitivity, I will keep that in mind as I test things. I should include the room/ambient temperature as part of my testing strategy when taking measurements.

Thanks for all the help/advice

 

[Karadjgne]

I'll be so happy when my parts come in and I can start twiddling with all this lol. Covid shipping got me messed up.

Room ambient really don't affect loads to any major degree, they might if the room was 40°C +, but normal room temps are so far behind what wattage the cpu can dish out under loads, that it doesn't really matter.

Idle is a different story. Cpu has such low wattage outputs that ambient temp affects interior case temps, which affects the cooler. It's physically impossible to cool a cpu lower than ambient temps by means of mechanical cooling. Both air and liquid cooling are mechanical. So if the room is 22°C, depending on the efficiency of the cooler, and the cpu, the case will be 6-12°C hotter. The coolers ambient air is the air inside the case, so you'd be looking at minimums of @ 30 - 36°C ish possible.

If room ambient is 32°C, idles will be 42 - 48°C etc. Of course Ryzen idle will mess all that up, the numbers anyway, not the actual properties or theory. It's not unheard of for Ryzens to idle at @ 60°C and Prime95 at 70°C ish.