Question Undervolting vs Power limiting the i5-13600k

NiceGuy1234

Reputable
Dec 26, 2019
9
0
4,510
I'm interested in tuning the i5-13600k to a bit more power efficient, but from what I've read is that 50/50 that turning on a power limit of like 90w doesn't impact performance on processor. Currently using an Asus Prime-k D4 b660 motherboard. Would it just be better to just set a power limit and call it a day rather than stress testing it for hours on end to make sure it's stable?
 
CPUs and GPUs operate on a frequency/voltage graph/lookup table.
  • Undervolting offsets that curve (ie, 5GHz=1.2V is now 5GHz=1.1V) and therefore improves efficiency at EVERY* frequency. Undervolting alone does not decrease performance (if done correctly), in fact, it may even INCREASE performance by allowing the additional power headroom to afford higher frequencies.
  • Power limiting (depending on how much) can range in performance loss depending on how consistent the workload is and/or how many cores are being loaded. The other main point being that, these days, most CPUs/GPUs are pushed beyond their "optimal" efficiency range by Intel/AMD/Nvidia at stock operation. By limiting power, you cut off that "inefficient" band of frequencies. Generally (at least initially) you get more power savings than performance loss (ie 10% power reduction =< 10% performance loss)
    • This stuff was all pretty well demonstrated by the recent wave of AMD non-X 7000 series CPU launch reviews since those chips are all just power-limited versions of their "X" counterparts.
As far as which one you should do....just know that power limits do limit [potential] performance to some degree as explained above. The nice thing (as you mentioned) is that you don't need to stability test anything because you're not altering the frequency/voltage table.

I have my 5600G undervolted. It draws so little power at 100% power limit, it doesn't matter
I have my 3060Ti undervolted AND power limited (80% so 180W instead of 220W). This is (primarily) to keep it silent while gaming. I still get more than enough FPS for my needs.

* assuming you're applying a global voltage offset, which would adjust every frequency/voltage point by that same amount. Global offsets are the "easiest" since you're testing 1 input across a frequency band. But it may not be the MOST efficient you can achieve since it assumes the stock curve was simply offset equally above the optimal curve of your chip across the entire frequency band. Most people only care about efficiency at the highest frequencies (maybe 4GH-5GHz) so a global offset is typically sufficient**

**assuming that the offset you land on is stable at lower frequencies also. Otherwise you may need to get more granular if you want to preserve your "high frequency" offset value.
 
Last edited:
  • Like
Reactions: Roland Of Gilead

NiceGuy1234

Reputable
Dec 26, 2019
9
0
4,510
CPUs and GPUs operate on a frequency/voltage graph/lookup table.
  • Undervolting offsets that curve (ie, 5GHz=1.2V is now 5GHz=1.1V) and therefore improves efficiency at EVERY* frequency. Undervolting alone does not decrease performance (if done correctly), in fact, it may even INCREASE performance by allowing the additional power headroom to afford higher frequencies.
  • Power limiting (depending on how much) can range in performance loss depending on how consistent the workload is and/or how many cores are being loaded. The other main point being that, these days, most CPUs/GPUs are pushed beyond their "optimal" efficiency range by Intel/AMD/Nvidia at stock operation. By limiting power, you cut off that "inefficient" band of frequencies. Generally (at least initially) you get more power savings than performance loss (ie 10% power reduction =< 10% performance loss)
    • This stuff was all pretty well demonstrated by the recent wave of AMD non-X 7000 series CPU launch reviews since those chips are all just power-limited versions of their "X" counterparts.
As far as which one you should do....just know that power limits do limit [potential] performance to some degree as explained above. The nice thing (as you mentioned) is that you don't need to stability test anything because you're not altering the frequency/voltage table.

I have my 5600G undervolted. It draws so little power at 100% power limit, it doesn't matter
I have my 3060Ti undervolted AND power limited (80% so 180W instead of 220W). This is (primarily) to keep it silent while gaming. I still get more than enough FPS for my needs.

* assuming you're applying a global voltage offset, which would adjust every frequency/voltage point by that same amount.


Thank you for the explanation. I am thinking of at least giving undervolting a try, but I'm not sure what kinds of stress test should I put it through? Like do I leave the PC on doing y-cruncher for extended hours? Or do I just play games normally and see if they crash off of that?

And do you happen to have any good guides for undervolting? I heard there are a lot of tools without going to the BIOS like throttlestop or XTU but I'm not sure if they are reliable or not.
 
Thank you for the explanation. I am thinking of at least giving undervolting a try, but I'm not sure what kinds of stress test should I put it through? Like do I leave the PC on doing y-cruncher for extended hours? Or do I just play games normally and see if they crash off of that?

And do you happen to have any good guides for undervolting? I heard there are a lot of tools without going to the BIOS like throttlestop or XTU but I'm not sure if they are reliable or not.
For general stability checks, running something that's CPU intensive for at least 10 minutes is generally good enough. If you want peace of mind, sure, run something overnight or however long it takes to help you sleep better at night. I like to use Cinebench R23 because it's already a well known CPU benchmark and it runs for 10 minutes. And doing this gets to kill another bird: I can see how much the performance changed.

As far as undervolting goes, figure out where the setting is in your UEFI settings, then start small, like one or two steps. Go run a stress test for a few minutes as a sanity check as if there's a problem with the setting, it'll typically show up almost immediately. Keep doing this the system crashes, then dial it back to the last setting that worked, then run a longer stress test or just do what you normally do and see if it survives.

Keep in mind though that undervolting does bring the CPU closer towards the edge of instability, so if you start having weird problems later down the road, dial the settings back a little more.
 
  • Like
Reactions: NiceGuy1234

Phaaze88

Titan
Ambassador
Would it just be better to just set a power limit and call it a day rather than stress testing it for hours on end to make sure it's stable?
Bios updates can break undervolts/voltage offsets. If you stay on a platform long enough, you will end up updating bios for one reason or other, and some of those updates come with new, improved voltage curves.
That's when you'll have to do the trial and error process all over again, finding new settings.
 
Bios updates can break undervolts/voltage offsets. If you stay on a platform long enough, you will end up updating bios for one reason or other, and some of those updates come with new, improved voltage curves.
That's when you'll have to do the trial and error process all over again, finding new settings.
Generally speaking you're supposed to reset the UEFI settings to their defaults after ( Or maybe it's before) an update anyway, if the update mechanism doesn't do it for you.
 

Phaaze88

Titan
Ambassador
Generally speaking you're supposed to reset the UEFI settings to their defaults after ( Or maybe it's before) an update anyway, if the update mechanism doesn't do it for you.
Aye, the update mechanism is supposed to clear settings for you after updating bios, but it doesn't always work, so you do it manually.
Even after doing that, going back and reapplying the old voltage settings, they still may not be stable anymore.
The -0.090 offset I used to have on this old system stopped being stable after an update. I moved down to -0.050, and that worked for some time, until I updated bios again later down the line, and that stopped working too.
Now, I leave voltage on auto but have a 210w short and long power limit.
 
Software adjustments are nice because of their convenience in on-the-fly adjustments. Once I get a final setting though, I enter it in BIOS and abandon the software.

When undervolting/overclocking, I don't stress test any longer than 10 minutes at first. I do run for 2-12 hours once I've landed on my "final" settings though. Some of that depends on how many frequency points you are (or can) testing.

Depending how close you want to be to the razor's edge will determine how long you should test past that point. It's good to keep an overall view of what you're doing. I generally* tend to find the crash point and just add 20-30mV to that (10mV = 0.01V, so not much). Sometimes it's helpful to monitor CPU power draw before and after undervolting also. Just make sure you're keeping an eye out to make sure frequency is the same as your control test. When you get nitpicking on "small" values, it's a nice sanity check to just look at power draw and see your adjustments are only moving the needle a few watts.
 
Last edited: