Rtx 3000 series undervolt discussion

Its part of the guide in the op. Once you set your curve, youll want to set a power limit. I suggest between 80 and 90 but you should try it and see where you are happy with performance and Temps. At 90 you should basically be as fast as stock while running cooler. If you want even less power draw you can do 80 and sacrifice a few percent of performance for much lower power.
Ah okay. I need to find my sweet spots of V/Mhz curve (tried and tested by Time Spy and BMI) and then set power limit at 90% to witness the actual temperature change.
 
Ah okay. I need to find my sweet spots of V/Mhz curve (tried and tested by Time Spy and BMI) and then set power limit at 90% to witness the actual temperature change.
You can see in the table in the op when I set 80% power limit I was giving up at most 4% of performance, but saving 70 watts.
 
Yeah, it looks like it was a matter of the boost bin being a step higher. I changed the offset and now get the maximum frequency to land where I want it. Thanks.

I'm noticing that the curve I set gets reset after a while, though. As in, it goes back to the stock curve instead of the specific frequency/voltage offsets I put in. I'm usually folding when not gaming (which has helped identify good long term stability incidentally) but when I come back, inevitably the curve is back to normal. Looking at the frequency editor shows nothing has changed. I am using Afterburner 4.6.3 beta 3. Anybody else seen something similar?

That means you had a TDR, its where the driver crashes and recovers, it resets the GPU back to default. Your overclock isn't stable, back it off 15 or 30MHz and try that.
 
I think they knew AMD had competitive products this time around so they shipped the cards with more agressive clocks which in turn they had to loosen up the power levels for each frequency point because silicon quality varience.

There has always been 10-20% efficiency to be gained since atleast Maxwell, it's just more this time because GDDR6X.
Ya, I agree with that. I suppose there's a lot of subterfuge in the tech market. It would take just one engineer with a grudge to tip-off a competitor, even if receiving that tip is unethical at best. That Nvidia pushed the clocks this much means they were worried, and those of us under-volting are likely running the GPU near the originally designed specification.
 
What are my options if I’m out of time (or lazy!) and just want to set one generic lower voltage? Like getting a fast feeling for whether my chip works great with a lower voltage or not?

I wish there was an app for that, sigh.We have all these semi-auto overclocking apps, I want an auto under volt tester! :rolleyes:
 
For those that went through the time to do this... is it worth it? Do you see any noticeable gains? How long did the whole process take until you were satisfied with the results?
Still OCing my RAM, getting burnt out of the whole OC process, and if there's not much to be gained I'd rather not go through the trouble personally
 
Last edited by a moderator:
What are my options if I’m out of time (or lazy!) and just want to set one generic lower voltage? Like getting a fast feeling for whether my chip works great with a lower voltage or not?

I wish there was an app for that, sigh.We have all these semi-auto overclocking apps, I want an auto under volt tester! :rolleyes:
For those that went through the time to do this... is it worth it? Do you see any noticeable gains? How long did the whole process take until you were satisfied with the results?
Still OCing my RAM, getting burnt out of the whole OC process, and if there's not much to be gained I'd rather not go through the trouble personally
If you look at the table in the op, the results should speak for themselves. With my 3090, I got 100% of stock performance while using 35 less watts. Or if I sacrifice 4% of performance, I can save 20% power which is 70 watts. If you don't have time to thoroughly test, and aren't doing anything mission critical with the GPU, go ahead and do steps 1-6, should only take a few minutes. I picked the +105 baseline because I believe it will work fine for most people. Then go ahead and set your power limit, 90% if you want almost no sacrifice in performance to save 32-35 watts, or 80% if you want to give up a little performance to save 64-70 watts. Do whatever you do, play games, and if you don't experience a crash, you're good to load your profile on startup. If you get a crash, simply dial the offset down by 15MHz, but like I said, I think 105 is a pretty conservative starting point.
 
Hey thanks lots for your effort here! Would these values work for a 3080 as well? If so I will definitely try this quick and dirty way in a few days, and tweak more if it's painless enough :)
 
For those that went through the time to do this... is it worth it? Do you see any noticeable gains? How long did the whole process take until you were satisfied with the results?
Still OCing my RAM, getting burnt out of the whole OC process, and if there's not much to be gained I'd rather not go through the trouble personally
What "gain" are you talking about? Clock speed? This under-volt thread is not a "how much faster than stock can I run the GPU" exercise. That's the over-clock discussion taking place in other threads. This thread is about finding a curve that suits your needs. There's a few different ways to accomplish that, and none of them take very long to do. My goal was to find the sweet spot where the GPU boosts to a reasonable clock, while significantly lowering power consumption, and therefore heat, while remaining 100% stable. For my 3080, that is a 1845mhz boost clock at .850v.
 
What "gain" are you talking about? Clock speed? This under-volt thread is not a "how much faster than stock can I run the GPU" exercise. That's the over-clock discussion taking place in other threads. This thread is about finding a curve that suits your needs. There's a few different ways to accomplish that, and none of them take very long to do. My goal was to find the sweet spot where the GPU boosts to a reasonable clock, while significantly lowering power consumption, and therefore heat, while remaining 100% stable. For my 3080, that is a 1845mhz boost clock at .850v.
You can gain power efficiency, you can gain frames, you can gain lower thermals, you can gain lower noise levels. Gains. Gains aren't only over-clocking, you are correct.
 
Last edited by a moderator:
You can gain power efficiency, you can gain frames, you can gain lower thermals, you can gain lower noise levels. Gains. Gains aren't only over-clocking - thanks for saying the obvious and trying to sound superior just for the sake of it.
In my opinion, everything is better. You will get lower temps/noise/power draw at the same performance if you choose to set a power limit in line with stock performance (like I said, for me it was 90%). I've said it before and I'll say it again, you are NOT always power limited! If you are playing a game at a detail level or a resolution that does not cause the card to be power limited, you will be enjoying higher clock speeds/lower power draw than stock.
 
I have a "Nvidia 3060 Ti Founders Edition". Which UV Settings are good?
 
First of all, thank you for this Topic! I've been playing with my 3080 Suprim X ... was using increments of 95, but I will try to use 100 now and 105 and so on ... this board is super sensitive :D

Using Memory at +632

I'm not working towards lower power consumption but for max clocks.
 

Attachments

  • testing_1.png
    testing_1.png
    855.7 KB · Views: 0
  • testing_1_curve.png
    testing_1_curve.png
    19.9 KB · Views: 0
Last edited:
First of all, thank you for this Topic! I've been playing with my 3080 Suprim X ... was using increments of 95, but I will try to use 100 now and 105 and so on ... this board is super sensitive :D

Using Memory at +632

I'm not working towards lower power consumption but for max clocks.
you want to use a multiple of 15, so not 95 or 100. It would be 90, then 105. :)
 
Any specifics on why we should use multiples of 15? Sorry, it's really because I don't know :)

Thank you so much for all the input and knowledge base you got into this topic :)
 
Any specifics on why we should use multiples of 15? Sorry, it's really because I don't know :)

Thank you so much for all the input and knowledge base you got into this topic :)
because 15MHz is the amount between boost tables. If you set in between boost tables, it is going to just go to the one above or below it anyway, and will make it harder to isolate instability at a given frequency.
 
Ok, let us imagine I want to up my boost limit to... 2115.
How would be your process for that? Exactly the same but instead of not going above the limit boost on the default curve, keep going through the voltage axis until you reached the frequency you wanted?

Or simply go to the 1.1v point ... add the offset needed to reach 2115 and then start from 0.800mV until the 1.1v?

Thank you once again, for the time :)
 
Ok, let us imagine I want to up my boost limit to... 2115.
How would be your process for that? Exactly the same but instead of not going above the limit boost on the default curve, keep going through the voltage axis until you reached the frequency you wanted?

Or simply go to the 1.1v point ... add the offset needed to reach 2115 and then start from 0.800mV until the 1.1v?

Thank you once again, for the time :)
You may not be able to set 2115 stable on a given card. But, the thing is you can set it at any voltage, but either it can be unstable, or just not hit those points due to power limit. So if you set it to the far right slider, it might be stable since that's the highest voltage, but it will rarely use that much voltage because of power draw interfering and keeping it lower. You can try just going up the next point over from stock boost and seeing where it falls apart though, but that is definitely harder because if you set something on the edge of where it gets power limited, when you test for stability it will not always be hitting that point and you might think it's stable just because it's not getting up that high.
 
mnewxcv In my case, when I set +105 at the relevant points, I'll end up hitting the stock max point (2070MHz) at 1000mv.
If I am aiming slightly higher, say +135, I'll get to 2025MHz at 950mv, but if I were set +135 at 975mv I'll be at 2055MHz.
Applying this would result in a bump in the curve further on as there will be points on the graph higher than the point at +135 (2055) 975mv

If I were to set +135 at 1000mv I'd be over the stock point. Should I do this, or add +150 to the 975mv point instead, bringing the 975mv point up to the stock 2070MHz?
 
You may not be able to set 2115 stable on a given card. But, the thing is you can set it at any voltage, but either it can be unstable, or just not hit those points due to power limit. So if you set it to the far right slider, it might be stable since that's the highest voltage, but it will rarely use that much voltage because of power draw interfering and keeping it lower. You can try just going up the next point over from stock boost and seeing where it falls apart though, but that is definitely harder because if you set something on the edge of where it gets power limited, when you test for stability it will not always be hitting that point and you might think it's stable just because it's not getting up that high.

Yes yes, 2115 was an example... just trying to understand how we used the same approach to get higher boosts OC other than just "limiting" the boost to the stock values.
 
Last edited:
Thanks for this easy to follow guide. I've applied this to my RTX 3060Ti and the results were very satisfying, much better than the auto OC scanner.
Although my card is power limited (the cheapest GALAX model), it's still using less power and lower temp, while still giving me a more consistent higher clock.

One thing with Timespy, on my last run that I spotted a problem, it still finished with no error and gave me a higher score which shouldn't have happened. Also I didn't have to wait 15 minutes between each run, since the last pass is a CPU test and gives the GPU a bit of time to cool down. Just needed to turn the fan a bit higher for a minute or so to get it back under 30 degrees.

Overall, well worth the time and very enjoyable process.
 
Thanks for this easy to follow guide. I've applied this to my RTX 3060Ti and the results were very satisfying, much better than the auto OC scanner.
Although my card is power limited (the cheapest GALAX model), it's still using less power and lower temp, while still giving me a more consistent higher clock.

One thing with Timespy, on my last run that I spotted a problem, it still finished with no error and gave me a higher score which shouldn't have happened. Also I didn't have to wait 15 minutes between each run, since the last pass is a CPU test and gives the GPU a bit of time to cool down. Just needed to turn the fan a bit higher for a minute or so to get it back under 30 degrees.

Overall, well worth the time and very enjoyable process.
welcome to the forum, thanks for posting!
 
I hadn't seen this method posted by anyone, so thought I'd share for anyone that hasn't seen it.

This will retain the original curve up to wherever you put in the horizontal section to limit clock speed.

First, reset back to stock in Afterburner,

Stock.PNG


Then apply a positive offset (I did +150 here),

Offset.PNG


Then, hold down Shift-Alt and select all of the area covering all of the pts to the right of your target voltage (I selected everything above 900 mV here), and then drag any of the points in this area down below your target frequency,

Moved_Pts.PNG


And hit apply,

Final.PNG
 
Thanks for the shortcut to drag multiple points, but doing it your way, you end up increasing the voltage for all frequencies below your target.

I think there is no perfect method... Your method is basically the opposite of wccftech method, since they lowered the entire curve first, and then raised the right part. On the other hand, you raised the entire curve, and then lowered the right part. In both cases you end up with frequencies below your target to have higher or lower voltage than stock. There is also OP method of course, which I don't quite like, since you end up with a trapezoidal curve which isn't optimal.

I can't wrap my head around what to do once I get my 3070 in all honesty. I also can't understand why you should want to lower power limit to save a few Watts whereas most people want to flash a BIOS with higher power limit.
 
70 watts, even assuming you put 100% load on the GPU for 8 hours a day - you are talking about $25 in savings over a year. Not worth the hassle IMO:

View attachment 310914

View attachment 310915

https://www.rapidtables.com/calc/electric/electricity-calculator.html
For me, it isn't so much the cost savings with electricity as it is:
  • More watts equals more heat, which usually results in lower average clock speed and/or higher fan speeds, thus more noise.
  • Applying more voltage than necessary to reach a certain clock speed when that same clock speed will run stable at a lower voltage is wasteful and, again, means more unnecessary heat and fan noise.
So it is more for optimizing than savings. My GPU sits idle more hours of the day than it is running at full blast, so cost savings isn't really going to benefit me much.
 
Thanks for the shortcut to drag multiple points, but doing it your way, you end up increasing the voltage for all frequencies below your target.

I think there is no perfect method... Your method is basically the opposite of wccftech method, since they lowered the entire curve first, and then raised the right part. On the other hand, you raised the entire curve, and then lowered the right part. In both cases you end up with frequencies below your target to have higher or lower voltage than stock. There is also OP method of course, which I don't quite like, since you end up with a trapezoidal curve which isn't optimal.

I can't wrap my head around what to do once I get my 3070 in all honesty. I also can't understand why you should want to lower power limit to save a few Watts whereas most people want to flash a BIOS with higher power limit.

Well you get the stock curve below target with the same increased offset as the target. I have found that this is close to optimal for my card. In the 875 to 925mv range, +150 is about all it can do. The only time this comes in to play is when you set the target high enough that the power limit is hit some/all the time and the GPU is running up/down the curve. And when it does this, it is almost always in the 875 to 925 range. Therefore it is optimal and a lot easier to create the curve than moving individual pts.

Different wat to get to the same place, but I found this method to be faster which is handy while testing.
 
Well you get the stock curve below target with the same increased offset as the target. I have found that this is close to optimal for my card. In the 875 to 925mv range, +150 is about all it can do. The only time this comes in to play is when you set the target high enough that the power limit is hit some/all the time and the GPU is running up/down the curve. And when it does this, it is almost always in the 875 to 925 range. Therefore it is optimal and a lot easier to create the curve than moving individual pts.

Different wat to get to the same place, but I found this method to be faster which is handy while testing.
What GPU do you have btw? I think I'll try to get the lowest voltage I can while keeping the 2040MHz boost on my 3070.
 
What GPU do you have btw? I think I'll try to get the lowest voltage I can while keeping the 2040MHz boost on my 3070.
It's a Zotac 3090 Trinity. It does have the Gigabyte bios on it too in order to raise the power limit to 390w. That's relevant to you question because that bios is 1755mhz stock. Hence 1755+150. May not be that relevant to your card. Not sure how the 3070 curve compares.
 
It's a Zotac 3090 Trinity. It does have the Gigabyte bios on it too in order to raise the power limit to 390w. That's relevant to you question because that bios is 1755mhz stock. Hence 1755+150. May not be that relevant to your card. Not sure how the 3070 curve compares.
What was your stock voltage when boosting to max stock frequency, and how did you choose to lower it to 900mV?
 
70 watts, even assuming you put 100% load on the GPU for 8 hours a day - you are talking about $25 in savings over a year. Not worth the hassle IMO:

View attachment 310914

View attachment 310915

https://www.rapidtables.com/calc/electric/electricity-calculator.html
I won't be doing it for savings, but rather to lower the heat being poured out into my room as well as potentially a better OC for my 3080 FTW3 Hybrid once it arrives. You're missing the point entirely.
 
What was your stock voltage when boosting to max stock frequency, and how did you choose to lower it to 900mV?
Well there's not one answer to the first question. It depends on what you are running. A game like Cyberpunk with heavy use of RT will result in a lower max frequency and voltage because the GPU will pull more power at a given frequency/voltage. The 900mV is about where the card runs out of power in Cyberpunk (390w). OTOH MSFS2020 will go up to 925mV or even a little more in 4k Ultra despite it only getting 30-50 fps.

I think for max performance, I'm actually going with a maximum of 950mV. This will let most games run at the power limit while keeping the card from boosting really high at low load which can cause a crash at +150 core clock. Going back to the previous paragraph, the GPU will boost all the way to max voltage at a low load sometimes since it isn't near the power limit.
 
I hadn't seen this method posted by anyone, so thought I'd share for anyone that hasn't seen it.

SNIP

Hmm, this method seems interesting. I have done a couple of quick comparisons, running the Time Spy stress test and comparing average frequency and temperatures, with the curves set at 32c, using my FE's stock fan curve. Target point is 1980@925mv
A run with the pull the whole curve down then pull the voltage point you want up to the frequency you want to target, and applying that, causing a steep climb prior to the point, then a flat line.
A run with the pull the whole curve up as such that the mv/frequency you want to target aligns, and then selecting all following points and pulling them down, and applying that, resulting in a 'stock but at higher frequency' curve, with flat line after the desired point.

Both runs result in identical average frequency, 1951MHz, and temperature, 74c. As this frequency and voltage combination do not exceed the extended power limit, the only factor that will reduce frequency will be thermals, which are also kept in check.

But, should the card throttle due to hitting thermals, or if some game/test manages to allow the card to hit the power limit, the frequency steps down the boost curve are larger when using the first method, while with the second method the frequency steps down are smaller, which should result in a smaller loss in performance.

But as I aim to keep my card from hitting the power limit, and the stock fan curve doesn't generally allow my card to exceed 75c, both methods result in exactly the same performance for me.

If you are targetting a higher frequency/voltage and are hitting power/thermal limits, the second method should in theory result in less loss of performance as the card won't step down in frequency as abruptly.

The following are the first few steps my boost curve will follow should it throttle for any reason using the two above methods

Method 1:
1980@925mv
1935@918mv
1890@912mv
1845@906mv
1800@900mv

Method 2:
1980@925mv
1965@918mv
1950@912mv
1950@906mv
1935@900mv
1920@893mv

Method 2 sees four steps down to reach the same frequency seen in just one drop in Method 1. I'd hazard a guess that Method 1 would be more stable when throttling as the frequency vs voltage points are lower. Needs testing!
 
If the target point is stable, it's likely the lower points will be also. I've found that as voltage increases, the highest stable offset decreases.

Which for me is really the point of this method. I found last night MSFS isn't stable at 950mv with a +150 offset. I have to cut the curve off at 925mV. This doesn't actually hurt performance as the only time the GPU goes above 925mV is with a low load (low power usage).
 
I have a multi point curve with an upper limit of 2070@975mv. As the multiple points result in larger drops down from 2070MHz as the card gets hotter/hits power limit, it remains stable.

Using your method, which I do prefer as it is essentially just shifting the stock curve upwards but clipping it to prevent the card boosting too far/capping the voltage, I have found that as the points directly below are less extreme, the card remains closer to this target for longer, but is stepping down the voltage, which I am finding isn't stable. For me, the next voltage point below is the same frequency, iirc 961mv.

This run is mainly just to see how far my card will go while remaining stable, but long term I'll be using the 1980@925mv cap as this combination keeps my card from exceeding the power limit.

Cheers for the idea!
 
So, I've just got my 3070, but can't see its voltage in MSI Afterburner. Is that because I'm using 4.6.2 instead of 4.6.3 Beta 2?
Besides, I did one run of Heaven at stock and noticed the GPU was always power-limited. Is that normal?
 
Back
Top