As I began the process of OCing my 3080Ti, I was searching the forums to seek out good starting points for an undervolt. I was shocked reading the commentary in some of these threads. Not only were people still asking why in 2021, but some were flat out critiquing the "decision" to undervolt top end GPUs. I decided to chronicle my early attempts and make a new thread. Undervolting a GPU is the best way to MAXIMIZE its 24/7 performance, and has been for quite some time.
 
Lets start off with the baseline. Stock out of the box, my 3080Ti will boost to a maximum of 1980mhz and usually does so using 1.081V. It does not stay there long as boost does its thing. Running the Port Royal stress test for a total of 6 loops (11mins or so) was long enough to level everything off, and the stock results looked like this (click links if the forum makes the images tiny like it always does):
 
https://i.imgur.com/CjUbfVM.gif

 
Pretty typical. PerfCap with Pwr literally the entire time. And as you can see, that 1980mhz was severely diminished and ended up averaging 1830mhz during this run.
 
Now for the absolute maximum this GPU is capable of on air. I can run +175/1250 stable. I maxed the power limiter, cranked up the fans to 100%, and opened my chasis door for the absolute best case scenario. With these settings, the maximum boost is 2145mhz at the same 1.081V. Again, it does not stick around for long and the results end up looking like this:
 
 https://i.imgur.com/CTETKUX.gif


Again PerfCap with Pwr for the entire run, and the average boost ended up being 1981mhz down from the maximum of 2145. That's still pretty good, the average here is higher than the maximum at stock and it tracks pretty closely with the core adjustment. But keep in mind that is with the jet engine fans and the chasis door opened, like I said its just showing the best case scenario without going under water. 
 
Now for the undervolt. The idea here is to attempt to match your traditional "best case scenario" with settings that are more suitable to being run 24/7. Nobody is running 100% fan with the door open, so obviously that means I am now back on autofan and closed the chasis up. After some trial and error, my voltage curve ends up targeting 1980mhz at 0.912V. That is the maximum clock at stock, and the average clock in my "best case scenario" at almost 0.2V less. It looks like this:
 
https://i.imgur.com/NnTEnSv.jpg

 
Before I drag the point I am targeting, I like to reduce the core by around ~200mhz to ensure stability at the low end of the curve. Ive always found that if you just drag the curve up from a higher baseline, that you often have to tweak all points below the target manually or risk instability with random PerfCaps. Seldom worth the effort as you will spend 99% of the time locked at your target clocks with a proper undervolt. As evidenced by the following results:
 
https://i.imgur.com/JTszhae.gif

 
As you can see, that 1980mhz maximum boost is averaging 1967mhz. That is a substantial overclock. The clocks are locked at the maximum a good portion of the time, and will only drop down one tick to 1965mhz during the heavier parts (which I can basically remove by running 100% fan and opening the chasis again... but that obviously defeats the purpose of what I am trying to show). Minimal fluctuation. No PerfCap to be seen. Thermals equal to stock with a maxed power limiter and big time OC (and significantly cooler in scenarios where you aren't cranking 100% GPU usage the entire time). Lower voltage. And most importantly, performance that matches the "best case scenario" with all the aforementioned benefits. Performance that never changes, even when running the full loop for extended periods: https://www.3dmark.com/3dm/66085354
 
Undervolting is how you overclock GPUs with nVidia boost.

While I'm not a fan of locking voltage for everyday use and prefer reducing power limit plus an overclock only if you're thermally limited, I'd like to point out that the 3080 Ti FTW3 under reports power use by about 30w at the stock power limit, so it's actually pulling about 425w when running at 100% power.

Are you using Precision X1 to unvolt your GPU, if so can you give us a screen shot of X1?

Can you do undervoltage with X1?

How has your experience been using x1 to do a curve?

I was considering starting another thread on this very topic as well, but since this one is so fresh, and is very much the exact same topic, here is some more info, simply to really compound what Celeras has already stated, and shown you guys.
 
With my 3090 I knew exactly what I was getting into, I knew that this is a 400w card and needs the airflow to keep it happy, this is not an issue for me, I have plenty of air exchange. What I do like however, is a quiet computer. I am very happy with the noise level of the 3090 FTW3 Ultra @ 70% fan speed, it falls in nicely with the 7 case fans all running around 50% duty. The total system noise sits right where I am happy. Though the 3090 however with extended sessions, sits hotter then I want. Now many people would say these temps are perfectly fine, I am now the normal person, I like stuff to be cool. So I decided to see how low the voltage could be offset and still hit evgas advertised 1800mhz on the 3090, for this card the voltage floor was 825mv.
 
I've been running this card like this for over a month now, and I am ever so pleased with the performance, and thermals, and this post drove me to finally log the differences to share with others. So, thanks Celeras for your post.
 
Now for the results!
 
Port Royal looping was used just like with the OP.
 
Out of the Box 3090 FTW3 Ultra, Fans locked at 70%

 
Stock performance voltage over time.

 
Now, same test, but boost curve modified to target 1800mhz @ 825mv

 
The raw data:
 
3090 Stock 1045mv-950mv
Max Power Draw 430.6w
Avg Power Draw 416.0w
System Power Draw 606w
Max GPU Temp 83.3°C
Max Hotspot 96.8°C
Max Memory 90.0°C
 
3090 1800 @ 825mv
Max Power Draw 317.8w
Avg Power Draw 293.57w
System Power Draw 448w
Max GPU Temp 62.5°C
Max Hotspot 73.3°C
Max Memory 76.0°C

Max Power decrease 26.2%
Avg Power decrease 29.4%
Sytem Power decrease 26%
Stock Hot Clockspeed 1830mhz
Undervolt Clockspeed 1830mhz

andross182
OP has some silicon lottery black magic with a +175 on the core...

yeah I have never seen an air-cooled 3080 ti hit any where near 2145 mhz. most people are lucky to even hit 2045 for more than a few seconds. 

Celeras
As I began the process of OCing my 3080Ti, I was searching the forums to seek out good starting points for an undervolt. I was shocked reading the commentary in some of these threads. Not only were people still asking why in 2021, but some were flat out critiquing the "decision" to undervolt top end GPUs. I decided to chronicle my early attempts and make a new thread. Undervolting a GPU is the best way to MAXIMIZE its 24/7 performance, and has been for quite some time.
 
Lets start off with the baseline. Stock out of the box, my 3080Ti will boost to a maximum of 1980mhz and usually does so using 1.081V. It does not stay there long as boost does its thing. Running the Port Royal stress test for a total of 6 loops (11mins or so) was long enough to level everything off, and the stock results looked like this (click links if the forum makes the images tiny like it always does):
 
https://i.imgur.com/CjUbfVM.gif

 
Pretty typical. PerfCap with Pwr literally the entire time. And as you can see, that 1980mhz was severely diminished and ended up averaging 1830mhz during this run.
 
Now for the absolute maximum this GPU is capable of on air. I can run +175/1250 stable. I maxed the power limiter, cranked up the fans to 100%, and opened my chasis door for the absolute best case scenario. With these settings, the maximum boost is 2145mhz at the same 1.081V. Again, it does not stick around for long and the results end up looking like this:
 
 https://i.imgur.com/CTETKUX.gif


Again PerfCap with Pwr for the entire run, and the average boost ended up being 1981mhz down from the maximum of 2145. That's still pretty good, the average here is higher than the maximum at stock and it tracks pretty closely with the core adjustment. But keep in mind that is with the jet engine fans and the chasis door opened, like I said its just showing the best case scenario without going under water. 
 
Now for the undervolt. The idea here is to attempt to match your traditional "best case scenario" with settings that are more suitable to being run 24/7. Nobody is running 100% fan with the door open, so obviously that means I am now back on autofan and closed the chasis up. After some trial and error, my voltage curve ends up targeting 1980mhz at 0.912V. That is the maximum clock at stock, and the average clock in my "best case scenario" at almost 0.2V less. It looks like this:
 
https://i.imgur.com/NnTEnSv.jpg

 
Before I drag the point I am targeting, I like to reduce the core by around ~200mhz to ensure stability at the low end of the curve. Ive always found that if you just drag the curve up from a higher baseline, that you often have to tweak all points below the target manually or risk instability with random PerfCaps. Seldom worth the effort as you will spend 99% of the time locked at your target clocks with a proper undervolt. As evidenced by the following results:
 
https://i.imgur.com/JTszhae.gif

 
As you can see, that 1980mhz maximum boost is averaging 1967mhz. That is a substantial overclock. The clocks are locked at the maximum a good portion of the time, and will only drop down one tick to 1965mhz during the heavier parts (which I can basically remove by running 100% fan and opening the chasis again... but that obviously defeats the purpose of what I am trying to show). Minimal fluctuation. No PerfCap to be seen. Thermals equal to stock with a maxed power limiter and big time OC (and significantly cooler in scenarios where you aren't cranking 100% GPU usage the entire time). Lower voltage. And most importantly, performance that matches the "best case scenario" with all the aforementioned benefits. Performance that never changes, even when running the full loop for extended periods: https://www.3dmark.com/3dm/66085354
 
Undervolting is how you overclock GPUs with nVidia boost.

Thanks, do you have some 3DMark Timespy GPU and Port Royal scores at these settings?

Lordred
I was considering starting another thread on this very topic as well, but since this one is so fresh, and is very much the exact same topic, here is some more info, simply to really compound what Celeras has already stated, and shown you guys.
 
With my 3090 I knew exactly what I was getting into, I knew that this is a 400w card and needs the airflow to keep it happy, this is not an issue for me, I have plenty of air exchange. What I do like however, is a quiet computer. I am very happy with the noise level of the 3090 FTW3 Ultra @ 70% fan speed, it falls in nicely with the 7 case fans all running around 50% duty. The total system noise sits right where I am happy. Though the 3090 however with extended sessions, sits hotter then I want. Now many people would say these temps are perfectly fine, I am now the normal person, I like stuff to be cool. So I decided to see how low the voltage could be offset and still hit evgas advertised 1800mhz on the 3090, for this card the voltage floor was 825mv.
 
I've been running this card like this for over a month now, and I am ever so pleased with the performance, and thermals, and this post drove me to finally log the differences to share with others. So, thanks Celeras for your post.
 
Now for the results!
 
Port Royal looping was used just like with the OP.
 
Out of the Box 3090 FTW3 Ultra, Fans locked at 70%

 
Stock performance voltage over time.

 
Now, same test, but boost curve modified to target 1800mhz @ 825mv

 
The raw data:
 
3090 Stock 1045mv-950mv
Max Power Draw 430.6w
Avg Power Draw 416.0w
System Power Draw 606w
Max GPU Temp 83.3°C
Max Hotspot 96.8°C
Max Memory 90.0°C
 
3090 1800 @ 825mv
Max Power Draw 317.8w
Avg Power Draw 293.57w
System Power Draw 448w
Max GPU Temp 62.5°C
Max Hotspot 73.3°C
Max Memory 76.0°C

Max Power decrease 26.2%
Avg Power decrease 29.4%
Sytem Power decrease 26%
Stock Hot Clockspeed 1830mhz
Undervolt Clockspeed 1830mhz

What did you use for the graphs

TenshiMarcelo
What did you use for the graphs

I logged the data, imported the csv into Excel and extrapolated the data into graphs.

Too bad the OP's method with the curve editor is the incorrect way to do it and will cause stutters in game due to the effective clock being lower with the curve points being so far apart.

Sultan.of.swing
Too bad the OP's method with the curve editor is the incorrect way to do it and will cause stutters in game due to the effective clock being lower with the curve points being so far apart.

Wrong until proven.
 https://www.hizliresim.com/6atik0r
My curve with 3090 xc3 ultra looks like this. I give +100 to core and flatten the chart at 1920/931mv. It is stable at any game and test I tried so far.

Sultan.of.swing

Bekcak

Sultan.of.swing
Too bad the OP's method with the curve editor is the incorrect way to do it and will cause stutters in game due to the effective clock being lower with the curve points being so far apart.

Wrong until proven.
 https://www.hizliresim.com/6atik0r
My curve with 3090 xc3 ultra looks like this. I give +100 to core and flatten the chart at 1920/931mv. It is stable at any game and test I tried so far.

Well did you even look at your curve compared to the OP's curve?  I'm going to assume you didn't.
Your curve is completely normal and is the way the curve should be setup which is why its fine, If you do not know what causes the effective clock to be incorrect with the curve editor set incorrectly I'd suggest doing some research.

 
@Sultan.of.swing
 
It would be more helpful to the community, if you shared specifically what you see as wrong based on your experience - or - a better way to do it.
 
I also believe that people should do some web searching, read what others have done = do your homework & then ask questions, to see if they comprehend the intricacies of any mod.
 
However, just saying something is wrong without good guidance of what & why; can be viewed as argumentative & does not foster communication or education