cancel
Showing results for 
Search instead for 
Did you mean: 

ROCm Discussions

eikon
Adept I

7950x build - Idle, Eco Mode, Undervolting

Hey all.  First off, apologies, this will be a long thread as i have many many thoughts 🙂
 
Its been 3yrs since iv posted here.  I joined when i built my 3950x system, and, coming from an Intel back then, wanted to try and control the temps on the AMD as they were much warmer.  Im back in the same boat now with my new 7950x build.  I know its not news that these new chips run quite warm/hot, and ill admit i know very little about whats going on under the hood, and how to optimize things in BIOS, but iv been fiddling around with settings for a few days, watching videos, trying to get my head around things, and a couple points have me confused.
 
First off, build info for this workstation:
7950x
MSI Meg Ace x670e
64gb Fury 5200 (xmp not enabled currently)
Liquid Freeze 2 280 AIO / Noctua HT2 paste
1000w seasonic
MSI gaming 3060ti - display
Asus 3090 - rendering
 
The 1st thing i dont understand is the voltage on these new chips.  When i run HWinfo, the stock settings dont drop the voltage below 1.39v at idle.  On my previous 3950x system, the voltage would drop down to 0.8xx V at idle (and i was using a mild -0.1v (iirc) entered as an undervolt), and the system was stable.   I was expecting things to operate similarly on this 7950x, but perhaps im wrong?  Surely having the voltage constantly sitting at 1.3xx is leading to increased temperatures or power consumption?  Or is that not how it works anymore?
 
After some youtubing, and trying to get my head around the new Eco Mode, settings that i could change and understand, i tried various configurations.  Nothing will allow the voltage to drop as low as i was expecting/hoping.  Again, maybe this is just me being misinformed... or wishful.
 
Across the board for my tests, my idling temperature never goes below 39C.  I was expecting it more in the mid 30s, but again, perhaps not how things work with the 7950.  This idle temp lower limit remains if i tried Eco mode, or setting a manual undervolt of up to (or down to?) -0.125v.  I was expecting that lowering the voltage by that amount would surely have an effect on the idle temps... but it didnt.  It idles at the same temp as stock.  So again, confused.
 
Since this is a workstation, im using Cinebench R23 as my stability test.  I had quite a few batches of settings that crashed while windows was loading, arriving at the desktop, or during my 1st attempt to write this post.  As of this moment, things seem stable, but with undervolting, it seems the instability happens at lower loads, not higher loads.  I was getting better CB23 results with undervolting than with Eco Mode.  My admittedly uneducated guess is that Eco mode is doing nothing for the idle voltage, and only adjusting the power under load, not allowing it to go too crazy.  So while everyone says Eco mode is great (at it def helps with load temps), it wasnt what i expected it would be.  Here are some results...
 
Stock: 37200.  PPT 208w.  95C.
PBO Eco Mode: 35500.  PPT 140. 75C.
Undervolt (PBO -250mhz, -0.122V offset): 38000, PPT 180. 85C (PBO thermal limit)
 
(HWinfo never tells me im thermal throttling, but isnt that what the stock setting is, and my last setting there for 85C?  isnt that a thermal throttle?)
 
That final undervolt setting is where im at currently.  My understanding of the PPT #s i was getting is that this is the max power consumed during the tests?  My current PPT in HWinfo is around 27w.  So, im guessing this is the idle power consumption? 
 
Either way, as for the CB23 results above, the last wasnt my best score, but since the best score ultimately crashed as i was trying to write this, i discarded those results (38170).  My initial testing was done by setting the cores to fixed values, as i think is the standard way things are done?  But when i found you could shave off mHz in PBO, that seemed like a better way to do it as im guessing it still allows the system to boost each core to whatever it can achieve (rather than giving each core a fixed value), and I can just back that off a little to play a bit more safely with the undervolt.  So thats my thinking there.  No idea if its correct.  HWinfo currently tells me my core clocks are around 5.5ghz, with a min of 4.8.  I recall when i tested by setting all core to 5ghz, that it stayed there and never changed, and i guess thats the point.  So perhaps by doing it the PBO way with negative mHz, its allowing for the system to be more adaptable and more efficient?  I have no idea.
 
So for these current settings, HWinfo was telling me that, during the render, the core clocks went from 5.5gh, down to closer to 5 even.  Similarly with the voltage, its idling at 1.3v, but then during the render test, it drops down to about 1.09v.  Just curious why this is?  Why is it telling me my idle speed and voltage is higher than when the actual work is being done?
 
I briefly tried the curve optimizer as the ONLY adjustment i was running, but i didnt have much luck and couldnt even get into windows.  I just did this following the suggestions of a youtube video, which said PBO limits should be off when using the curve.  I dont know if thats true or not?  He was using the older 59xx chip.  Either way, I was most likely using too high of a setting (started at 20, worked my way down to 10, but still no luck), so i may revisit that idea later. 
 
So while there are a handful of questions throughout here, the main question is why am i stuck idling at 1.3v?  Does anything iv said above make sense?  Am i making reasonable adjustments or am I way off?
 
The final point, and somewhat of an aside to the above, is on my 280 cooler.  Is my 40C temp lower limit due to this?  When researching build components, i decided on the 280 vs the 360 because id seen a test that they were fairly similar in cooling ability.  The 280 has a higher air flow, but less surface area, whereas the 360 is the opposite.  Since the 280 is smaller overall and would allow me another fan up at the top of the case, i opted for that.  Im not sure how much (if any) improvement i would see with a 360.  All the above testing was done with the CPU fan on stock settings.  It wouldnt realy ramp up much during my tests reaching 85c.  But just now for kicks, i turned it up to 100% speed, idle temps remain at 39/40, and my CB23 score and temp remained the same.  Interesting though that nothing changed from full speed to like 50% speed. So perhaps if 40C is the lower limit of this cooler, the 360 might be a bit better since it has more surface area, as it seems the airflow isnt realy changing anything.
 
Also,  FWIW, perhaps iv been stingy on the thermal paste as well?  I realy didnt want any leaking out and getting underneath the IHS, so i probably purposely tried to keep it on the lighter side.  Temps had initially been higher, but it lead me to suspect something, so i removed the cooler and found that the paste had somehow only covered 1 side of the CPU, with very little on the other side.  After reapplying, its better now at my current 40c.  Im also unsure about overall screw tension when tightening down the cooler, but i thought it was better to err on the side of caution, and only tightened it until the screwdriver started to give me moderate resistance.  Definitely tighter than finger tight though.
 
Apologies for the long thread.  I just like to be as detailed as possible and try and include as much info as possible.  Any and all insights are welcome.  Thanks!
0 Likes
8 Replies
MADZyren
Paragon

Go to BIOS and change voltage curve to negative 10 at first. In short testing, I could go to negative 20. This improved performance, lowered temps and of course voltage.

I get >38000 points in CbR23 with slow 5200MT/s memory and no other changes than negative voltage curve.

Right amount of thermal paste depends on how smooth surfaces you have and how flat they are. When in doubt, having too much is better than not enough. Extra will simply squeeze out.

0 Likes

Im having 0 luck with the curve.  I couldnt even get in to windows at -7, nevermind -10.  Are there any other settings that i need to be using in conjunction with the curve?  Does PBO limits need to be on?  off?  motherboard?

0 Likes
eikon
Adept I

UPDATE!

So im not sure what the hell was going on originally.  I ended up wiping windows, doing a CMOS reset and starting over completely.  I knew i had seen my voltages drop to idle when i first started the system.  So something i did, or installed, must  have screwed that up.  

After starting fresh and paying close attn to what i was doing, i got everything installed and it now drops to idle at about 0.7xx after a min or so in windows.  So that seems normal.

I then had another try at both voltage offset and curve optimizer...

For some reason, this time i was able to get curve optimizer to work.  I was able to get it stable up to -15.  Once i started going beyond that it was unstable.  Im sure if i went in and did it per core i could improve it further, but im not quite sure how to do that yet, methodically.  I got a CB23 of 37500.  I got an almost identical score when i tried regular undervolting by -0.115v.  Anything lower than that and windows became unstable.

So while i watched HWinfo as CB was running, i was seeing that the 2 chips were pretty even.  One was faster, usually running around 5050ghz, and the other around 4950ghz on average.  HWinfo says the max boost is supposed to be 5500ghz, but i dont see that # reached at any time during the test.  I have HWinfo set on the default refresh rate of 2000, so maybe it happened so fast i missed it?  Either way, does this appear normal?

cb23_curve15_37500_xmpOFF.JPG

THis screenshot was from just after the test, with Curve -15, and XMP OFF, and a PBO thermal limit of 85 (though HWinfo still tells me im not thermal throttling, so i still dont understand that).  CB actually seemed to score better with XMP off (...?), and  I also noticed the idle temps were lower as well, by about 4c.

I came across a couple tooks called Boost Tester, and PBO2 Tuner which allow you to dial those in a bit in windows, but that video i found them in was using them on an older 5950x Zen3.  So im not sure if they work with my newer setup, and i didnt want to try it until i knew for sure.  Does anyone know if they can be used on Zen4?

0 Likes

These were my results using Noctua D15 air cooler. I think even more powerful cooler (water) might offer even better results.

I had a GPU installed. When I removed GPU and started using iGPU, I could no longer go anywhere this far. Still waiting to find a decent deal on a new GPU.

Cb23 : Win10 Pro, Asus B650E-E, 7950X, 2x32Gb Kingston (I believe it is Micron Dual-Rank) 5200 CL40 with Asus optimized memory profile (CL34)

Asterisk is the highest stable result I got. After that, while these tests were successful, further tests failed, but stability tested only using CB23 anyhow.

Curve optimizer with negative values : 1st run : 2nd run : 3rd run : In average:

Curve -237910379883787037923
Curve -438070378503806837996
Curve -638169382553828338236
Curve -838285379243844638218
Curve -1038242385493823538342
Curve -1238455380463859038364
Curve -1438520386353861038588
Curve -1638549383923866938537
Curve -1838515385733887538654
Curve -19*38644387403869538693
Curve -2038942389883875638895
Curve -2238814389233899438910
Curve -2438932383333883138699

 

0 Likes
eikon
Adept I

at -18 i was able to complete a run, but CB crashed at the end of the frame.  i tried to run it again, and windows crashed.  so i dont know what that score would have been.  looks like you got about 1000 pts higher than me across the board.  i guess you got a better CPU?  unless there are some other settings i have that have hindered my results.  I did some other small tweaks like LLC 4, disabling global c-state, as i heard those things help stability.

so ya, are your results above average or are mine below average?  ultimately i wont be doing workloads as intense as CB, but its nice to know where the upper limit is.  stability and temp/efficiency are my main concerns.

0 Likes

Actually AM5 MSI-boards have a BIOS issue, which will soon be fixed. It should improve performance. More information here: https://www.youtube.com/watch?v=UN2gkbMQ2fs&t=899s 

I might also have had slightly better luck with silicon lottery, but even at default setting I was already getting pretty good result in this benchmark. I wonder if it is because I use Windows 10 (although I have security features on, some of which are default only in Win11 and can degrade performance a bit) or is it because I have 64 gigs or RAM or because - to my knowledge, my memory is dual-rank unlike 2x16 Gb DDR5 kits.

 

0 Likes
eikon
Adept I

interesting.  there was a bios released a couple days ago, but i dont like to jump right into those without letting a few other people test them 1st.

in an odd turn of events, my thermal grizzly paste and cpu guards just showed up.  i turned off and cleaned off the old cpu paste (noctua).  despite my efforts to get a more even spread that time, it was better, but still lopsided to one side.  after applying the guard, i applied the kryonaut in an X shape as this seemed to give a realy good coverage in tests i saw.

strangly, the system booted up saying there had been a hardware configuration change?  not sure if it just does this anytime the CPU is removed?  cause nothing was actually changed.  anyway, i loaded my preset from this morning of the curve -15.  another oddity is that this MSI board only seems to have SOME of the settings you put int.  all the PBO stuff was not saved, so i re-entered it.

when i got back to windows, instead of the usual 45 seconds of loading time where i watch HWinfo and the voltage is somewhat up there, it dropped almost immediately to 0.9v.  i gave it its standard minute of loading windows stuff, and then it settled to its idle of 0.805v.  Not as low as it was going this morning, but now idling at a lower temp.  i saw it go down as 37.

i ran a CB just to see, and i got 100-300mhz more on some cores, but higher across the board.  It maxed at slightly higher voltage as well, which im guessing is what accounted for the higher clocks, but i dont know why this would be.  So either i missed some setting in bios that had actually been hurting me, or the Kryonaut paste just delivered some noticable gains.  I scored a 38300.  So now im wondering if i have the ability to go in a start adjusting the curve more and see if i can get to -20 maybe?  Anyway, interesting is all. 

As far as tuning the curve optimizer, if i have some cores that are NOT boosting as high as they should (which i assume is 5.5ghz based on what HWinfo is telling me), is this when i lower the curve on those cores?  like -15 to -18, or whatever?  Or do i go the other way, closer to 0?  Because in a video i watched, it seemed he would look for the cores that had effective clocks not getting as high as what they should be, and he would dial in a lower negative number, heading towards -30.  i know the more negative it gets, the less voltage its applying, but why is this having the effect of allowing the cores to boost higher?

0 Likes
eikon
Adept I

this thread doesnt seem to be getting much interest, but ill try this here 1st before starting another thread...

does anyone know if the zen3 tuning tools will work on this new generation?  I was watching this video...

https://youtu.be/wpciiIddJvE?t=196 

.. and hes using PBO2 Tuner, Boost Tester, and Core Cycler.  Im pretty sure i saw a more recent video using core cycler, so im guessing that one would be ok.  But, im hesitant to try and run the other 2 if they arent realy meant for working with this generation.  Basically i think iv reached my stable all-core limit with Curve Optimizer at -14, which i wasnt super happy with.  Im assuming i can eek out a bit more if i go in and do it per-core, but im not realy sure how to do per-core testing on a 7950.

Anyone?  Thanks.

0 Likes