Overclockers: Why You Should Undervolt Your Graphics Card.

[Denwagg]

Anyone that has a high end Nvidia Graphics 3080, 3080 ti, 3090, 3090 Ti. You really ought to download MSI Afterburner and Learn how to undervolt you gpu.

Below is a screen shot of my system. Everything in DCS is MAX 3440x1440, AH64D, Syria, Quickstart, Hot Start default miz

Its really easy to learn you just tweak a graph in Afterburner. and then test it for lockups. It will not hurt your computer or GPU in any way. My 3090 Ti went from 

core clock 1920 to 2010 Mhz

GPU power 445 to 335 Watts

Temp from 83 to 65 C

Voltage from 1.020 to .920 volts

Stable no crashes

 

Regards

 

Edited August 1, 2022 by Denwagg

[Bossco82]

Couldn't agree more. Running the stock Nvidia algorithm the voltages can go upto 1.1Mv which can make it unnecessarily toastie. 

Almost all Nvidia GPU from the 1080 to the 3090ti can do roughly 1950Mhz at 0.975Mv which will save you peak wattage and temps. Play around with it to find your sweet spot but that a good starting point. 

[nachomaga]

Sorry for my ignorance, but what's the point to undervolt the GPU? For run with less power and get lower temps to extend GPU life? My 3090 usually gets hot (anyway, normal for this card).

The Core clock going a bit up is noticiable in fps? I'm in VR and every frame counts

And another question, how do you get the core clock go up while power going down? Thanks!

[Denwagg]

When I undervolted My Clock speed went from 1920 to 2010 Mhz and temps went down from 83 to 65. So I gain performance and lowered the heat and wear on my gpu. Its a win win situation

Regards

 

 

[Hiob]

1 hour ago, nachomaga said:

Sorry for my ignorance, but what's the point to undervolt the GPU? For run with less power and get lower temps to extend GPU life? My 3090 usually gets hot (anyway, normal for this card).

The Core clock going a bit up is noticiable in fps? I'm in VR and every frame counts

And another question, how do you get the core clock go up while power going down? Thanks!

Because, more voltage gives more stability (in general) and nvidia has to account for the serial quality spread of the GPU they sell. Basically, they need to make sure, that every sub-par binned GPU is running without issues after beeing sold. Therefore their default Voltage/clock curve has a lot (A LOT) of safety margin.

I absolutely agree with everything said above. Personally I run my 3080 at 2040MHz/925mV as a save stable daily driver. But it will go up to 2130 @ 975mV when outside temperatures are not as high as they are now. (Water cooled)

Power Consumption is about 300W instead of 350W and Performance about +3% over stock settings.

(GPU clock speed gain in % almost equals fps gain in % in case you wonder)

Edit: The reason that the cooling is relevant here, is that it the amount of voltage you need to sustain a certain clock speed depends on temperatue. And the silicone lottery determins how energy efficient the gpu is in the first place. So from the same brand and type, one GPU may need 950 mV to push over 2 GHz at a given temperature and the next on may need only 925 mV. Afterburner is a great tool to safely optimize all those variables for your individual GPU (mainly with the curve editor).

Your GPU running cooler is a positive side effect, that may prolong its life. But you can of coure choose your trade off. More Efficiency/less power consumption vs. more performance.
Coming from the stock settings you can have both up to a certain point (so more performance AND less power consumption/temps) - but once you found the optimum - from there you can choose either way.

Edit: btw. on the title of the thread - in principle all of the above is not limited to high-end gpus. (or to nvidia). You can optimize a 3060 in the same way. The results may not be as significant, but there is no reason, not to try.

 

For every one who want to try this. When you run into stability problems, raise your voltage in 25mV steps as long as you stay below or at  1000mv. If still unstable, lower the clock until stable. Voltages above 1000mV are pointless and almost guaranteed to be inefficient. (In stock 3080s run between 1030 and 1075 mV...)

Edited August 3, 2022 by Hiob

[nachomaga]

Thanks for explanation, guys I will give a try.

[Drac]

6 hours ago, nachomaga said:

Thanks for explanation, guys I will give a try.

Since you are a VR user, I will await your findings.

[SkateZilla]

This Applies To All Current Gen and Previous 2 Gens of GPUS.

AMD And nVidia

Edited August 3, 2022 by SkateZilla

[BitMaster]

It's the vicious circle run the other way round.

Less-Volts=cooler=morestable=higherclocks

It's the smart way to run that damn circle.

[Burt]

What’s the recommended drop in your core -200 - -250 ? 

[Hiob]

12 hours ago, Burt said:

What’s the recommended drop in your core -200 - -250 ? 

You mean in voltage? What are you aiming for? Efficiency or Performance? Either way, stay below 1000mV. A good point to start is 925-950 mV. Find the highest frequency you can run with this voltage and reduce one step (or don’t if you don’t mind occasional crashes).

But its not as easy as dialing in a negative offset. You need to modify the curve. I would recommend one of the countless youtube videos on that matter. But in short: Open the curve editor in afterburner (little bar graph symbol next to the core clock offset). Use the offset to lower the highest point of the curve below the clock you are aiming for. Then click and drag the dot at your target voltage and raise it to your target frequency. When you click apply, the curve should be flat from this point on. It can be a bit fiddly to hit the right numbers and the freq can only be adjusted in steps of 15 Hz.

Some Notes:

For my card and cooling I usually go with 975mV for Performance and 925mV for efficiency approach (you can go lower than that, e.g. 875mV but then you will noticeably sacrifice performance).

The increments for voltage are smaller than 25mV. You could go 987mV for example. I'm just biased towards "round" numbers.

Even if you found your setup and locked your clock at 2010 MHz e.g., the actual frequency may go up or down a step depending on the temperature conditions.
E.g. I set my card to 2040MHz@925mV but when I start from a completely cold system it would usually clock 2055 MHz. When it gets really heat saturated on a hot day it may drop to 2025MHz.

The whole thing is a matter of trial and error. You can apply the changes to afterburner in real time, while you run Heaven Benchmark in a window. Note though, that something that is stable in Heaven isn't necessarily in DCS... and when it is stable in DCS it may crash in heavier workloads like Cyberpunk2077 e.g.
Every application loads the GPU differently. Heaven is a good approximation for DCS  though.

If you don't care for efficiency that much, you can stabilize a given clock by raising the voltage a step or two.

DISCLAIMER:

Some may have noticed, that I happen to repeat myself over this matter in this thread and a couple of others.

The thing is, that I firmly believe, that there are only gains and no losses in looking into this stuff. Therefore I'm advocating for it.
When I first became aware of GPU-overclocking and Afterburner, I thought it was a complicated matter for enthusiasts and went totally above my head.
Nothing could be further from the truth. It is dead easy. Worst thing that can happen, the gains are minuscule or you introduce instabilities when you overdo it. In this case, you can always convert back to default with the press of a button. But usually you will gain something. Be it higher clocks, cooler temps or less energy consumption (or a mix).

So there you go. Try it!

Edited August 11, 2022 by Hiob

[Burt]

Wow thanks ! 
 

I don’t know why MSi afterburner doesn’t display with OXR. I use to like having it running with SVR and real-time tweaking with my G2 off my head and watching my monitor display with msi-afb. I’ve  been trying to figure out why it doesn’t do that with OXR - the monitoring displays

Edited August 11, 2022 by Burt

[skypickle]

Could you be more specific about the settings in MSI after burner? What does your voltage/frequency curve look like (ctrl - F in Afterburner)

 

null

[stormrider2]

2 hours ago, skypickle said:

Could you be more specific about the settings in MSI after burner? What does your voltage/frequency curve look like (ctrl - F in Afterburner)

 

null

 

[BitMaster]

Looking at the river beds and the temperature gauge, 100w less is great but not enough.

We should be gaming with 50-100w and not 300-600w.

 

This house of cards is currently falling apart if you havent noticed yet

[Lange_666]

I'm having a problem understanding how to get the reference clock speeds "all" these video's talk about.
None of them makes that clear, they all drop de curve down -200 tot -300Mhz.
Then comes the part where at undervoltage point, they move up that point particular point to a certain frequency (2:30 in the video above) without exactly telling how they defined that.
What i don't understand is how the define that particular frequency. They get by with saying "what you want" "or your desired clock speed".
In some video's they refer to the specs of the card, then point to the clock and boost speeds but then set a value that is never near the clock or boost speeds they just talked about. So what's the point then pointing to those values if they are not used?
If i put all the video's i watched side by side (even for my particular card) they all set it to a different value, none of them near the clock or boost speed. Based on what then?
So how do i determine the clock speed that i need to move up on the new undervolting value?

Edited September 9, 2022 by Lange_666

[Hiob]

1 hour ago, Lange_666 said:

I'm having a problem understanding how to get the reference clock speeds "all" these video's talk about.
None of them makes that clear, they all drop de curve down -200 tot -300Mhz.
Then comes the part where at undervoltage point, they move up that point particular point to a certain frequency (2:30 in the video above) without exactly telling how they defined that.
What i don't understand is how the define that particular frequency. They get by with saying "what you want" "or your desired clock speed".
In some video's they refer to the specs of the card, then point to the clock and boost speeds but then set a value that is never near the clock or boost speeds they just talked about. So what's the point then pointing to those values if they are not used?
If i put all the video's i watched side by side (even for my particular card) they all set it to a different value, none of them near the clock or boost speed. Based on what then?
So how do i determine the clock speed that i need to move up on the new undervolting value?

 

When you leave your card on defaults and just monitor its behavior (clock speed, voltage, temps), you will notice that the clock fluctuate a lot, but usually close to or below 2000 MHz at best. The voltage on the other hand will fluctuate but usually sit well above 1030mV.

The reason for the fluctuation is, that with such high voltages you will often hit the boundaries of the envelope - either due to temp or max current.

One of the main goals of undervolting is to get into a more efficient part of the curve/envelope. The high voltage by default is there to stabilize the GPU (higher voltage makes flipping transistors faster and more reliable. By default there is a lot of unnecessary safety margin at the expense of clock speed.

A good starting point is the highest clock your card will reach in default settings, but I'm yet to see a card that wouldn't easily hit 2000 MHz.

There are no hard fixed points. Just some common values by experience. From there you decide the direction. More energy efficient/cooler or more clock speed.

 

Edit: The bold part is the answer to your question!

Edit2: Mind that the specific clock vs. voltage vs. stability can be very dependent on the application. A stable overclock/undervolt for DCS may crash in a different game or vice versa if you really push it to the limits. 

So I would either recommend to find your best settings and than dial it back one step each (+1 step on voltage and -1 step on clock), or to use different presets for different games (and outside temperatures. Stable with 20° Room temperature may crash at 25°C ambient temp).

Edited September 9, 2022 by Hiob

[EightyDuce]

Yeah, this has been a thing for a while now but hasn't gotten much traction, probably because many are stuck in the old ways of overclocking where "moar voltage =moar better". 

Example, my wife's 3060ti, best it could do stock was 1960 core on stock settings. With added voltage it could run 2035 but it would be hot as crap, suck power and would still throttle down to low 1900s

With undervolting I could get to lock in at 937mV, hit 1995 consistently and stayed at 2035 more often than not, while using close to 30w less and 8c lower temp and performed better (w/ a +1100 on mem OC). 

[edmuss]

Playing with the undervolt on my 3080ti (asus tuf oc), with maximum power limits set it will just about top 1920 when benchmarking but will be bouncing down to 1800 and even high 1700s quite regularly.

Set the clocks to 1860 @875mv and she sits constantly at 1860 with occasional drops to 1845. If I try to push anything higher on core clock it crashes heaven benchmark dramatically, even if pushing the voltage up a little; I guess it must be the silicon limit.

I've run the memory at +1000 without a problem and a quick test at +1500 was still showing performance uplift.

Any thoughts on why I can't push the core clock higher? Looking at the 3Dmark results a lot of people are getting another 3-400mhz more.

[Lange_666]

@ Hiob: Could you take a look at the video in stormrider2's post above and explain a bit more what the dude is trying to say. Start around 2:20 in the video.

I'm struggling with 2 phrases:

1: "Drag down the top of the curve to your desired frequency".
- How do i determine my desired frequency? My 3080Ti has a base clock of 1365 MHz and a boost clock of 1725 MHz. However, the presented curve in my Afterburner graph goes to 1975 MHz from around 1090mV (and is flat from then on to 1250mV). In the video, the curve is dragged down 162 MHz overall but that is based on what? I have seen video's go to -300, most are in between -200 and -250 but on how much the curve needs to be dragged down or what the reference point is is never explained.
Or do i need to bring the top of the curve down to my cards boost freq?

2:  "Line that up with the voltage you want to run it at and press confirm".
- Now all these video's use different values above the curve but how do they determine that value? That is never explained.
Also, if i look at the curves in all the video's after they "applied" their choice, most of them (but not all) have a max value that is a bit lower then the max value presented by the default max value but at a lower voltage. I don't have a problem with the voltage value but with the value they move the frequency up again.

Somehow i fail to understand how they determine the values they set.


 

[Hiob]

17 minutes ago, Lange_666 said:

@ Hiob: Could you take a look at the video in stormrider2's post above and explain a bit more what the dude is trying to say. Start around 2:20 in the video.

I'm struggling with 2 phrases:

1: "Drag down the top of the curve to your desired frequency".
- How do i determine my desired frequency? My 3080Ti has a base clock of 1365 MHz and a boost clock of 1725 MHz. However, the presented curve in my Afterburner graph goes to 1975 MHz from around 1090mV (and is flat from then on to 1250mV). In the video, the curve is dragged down 162 MHz overall but that is based on what? I have seen video's go to -300, most are in between -200 and -250 but on how much the curve needs to be dragged down or what the reference point is is never explained.
Or do i need to bring the top of the curve down to my cards boost freq?

2:  "Line that up with the voltage you want to run it at and press confirm".
- Now all these video's use different values above the curve but how do they determine that value? That is never explained.
Also, if i look at the curves in all the video's after they "applied" their choice, most of them (but not all) have a max value that is a bit lower then the max value presented by the default max value but at a lower voltage. I don't have a problem with the voltage value but with the value they move the frequency up again.

Somehow i fail to understand how they determine the values they set.


 

…because they are completely arbitrary. The only reason the curve is dragged down in the first place, is to make sure that the value you are aiming for will be the highest and the curve is a flat top from there.

We want achieve a constant clock. Try the following. Lower the whole curve via dialing in a negative offset so that its highest point is around 1920 MHz.

Than take the dot at 975 mV and raise it to 2010 MHz and hit apply. Voila, you now should have a raising graph up to 2010 MHz at 975mV and a flat top from there.

Try to play DCS with those settings and report back. If it crashes, reset to default and repeat the process with 30 MHz less. But it should be fine. If it is stable, you can try either reduce the voltage further (950 mV) or raise the clock. How much you can raise the clock depends strongly on your cooling. When in doubt, opt for stable clock and lower temps.

The curve editor is a bit twitchy sometimes. When you want to change something, it is often easier to start all over.

[edmuss]

11 minutes ago, Lange_666 said:

@ Hiob: Could you take a look at the video in stormrider2's post above and explain a bit more what the dude is trying to say. Start around 2:20 in the video.

I'm struggling with 2 phrases:

1: "Drag down the top of the curve to your desired frequency".
- How do i determine my desired frequency? My 3080Ti has a base clock of 1365 MHz and a boost clock of 1725 MHz. However, the presented curve in my Afterburner graph goes to 1975 MHz from around 1090mV (and is flat from then on to 1250mV). In the video, the curve is dragged down 162 MHz overall but that is based on what? I have seen video's go to -300, most are in between -200 and -250 but on how much the curve needs to be dragged down or what the reference point is is never explained.
Or do i need to bring the top of the curve down to my cards boost freq?

2:  "Line that up with the voltage you want to run it at and press confirm".
- Now all these video's use different values above the curve but how do they determine that value? That is never explained.
Also, if i look at the curves in all the video's after they "applied" their choice, most of them (but not all) have a max value that is a bit lower then the max value presented by the default max value but at a lower voltage. I don't have a problem with the voltage value but with the value they move the frequency up again.

Somehow i fail to understand how they determine the values they set.


 

I don't think it matters how much you initially lower the core clock in afterburner as long as the highest point in the curve ends up below your target clock speed.

Once the whole curve has been offset downwards, select the curve point that corresponds with the maximum voltage you want to use and then drag that point back up to the clock speed you're aiming for, hit apply and it should then flatten the curve after the voltage and ramp up linearly before the voltage.

Trial and error determines the sweet spot of performance by adjusting the clock speed and temperature by adjusting the voltage. In addition to this you can increase the memory clock for a healthy boost.

Remember to check your stability with a benchmark, unigine heaven is a pretty good comparison to DCS (I think) and temperatures with gpuZ.

 

[Hiob]

1 hour ago, edmuss said:

Playing with the undervolt on my 3080ti (asus tuf oc), with maximum power limits set it will just about top 1920 when benchmarking but will be bouncing down to 1800 and even high 1700s quite regularly.

Set the clocks to 1860 @875mv and she sits constantly at 1860 with occasional drops to 1845. If I try to push anything higher on core clock it crashes heaven benchmark dramatically, even if pushing the voltage up a little; I guess it must be the silicon limit.

I've run the memory at +1000 without a problem and a quick test at +1500 was still showing performance uplift.

Any thoughts on why I can't push the core clock higher? Looking at the 3Dmark results a lot of people are getting another 3-400mhz more.

The Ti’s maybe a bit different in their behavior, because they are already pretty much on the edge. Best advice for Ti‘s is to go for lower temps/voltage and keep the clocks. 875mV seem very low though. At what voltages operates the card in default?

From my experience, DCS is extremly sensitive to memory overclocking. Though I can overclock the VRAM +1000 MHz on my 3080 and successfully run 3DMark, DCs will crash at +300 already. That is barely noticeable. Therefore I skipped the Vram overclocking altogether.

8 minutes ago, edmuss said:

I don't think it matters how much you initially lower the core clock in afterburner as long as the highest point in the curve ends up below your target clock speed.

Once the whole curve has been offset downwards, select the curve point that corresponds with the maximum voltage you want to use and then drag that point back up to the clock speed you're aiming for, hit apply and it should then flatten the curve after the voltage and ramp up linearly before the voltage.

Trial and error determines the sweet spot of performance by adjusting the clock speed and temperature by adjusting the voltage. In addition to this you can increase the memory clock for a healthy boost.

Remember to check your stability with a benchmark, unigine heaven is a pretty good comparison to DCS (I think) and temperatures with gpuZ.

 

He nailed it!

Edited September 9, 2022 by Hiob

[edmuss]

Just now, Hiob said:

The TIs maybe a but different in theit behavior, because they are already pretty much on the edge. Best advice for ti‘s is to go for lower temps/voltage and keep the clocks. 875mV seem very low though. At what voltages operates the card in default?

From my experience, DCS is extremly sensitive to memory overclocking. Though I can overclock the VRAM +1000 MHz on my 3080 and successfully run 3DMark, DCs will crash at +300 already. That is barely noticeable. Therefore I skipped the Vram overclocking altogether.

It peaks at about 1030mv I think, I started at 825mv but it wasn't stable much above 1600mhz. I've run a few hours of DCS at +1000 memory without a problem, more testing required at higher though.

Temps wise during heaven it's peaking at mid 70s core with memory a couple of degrees hotter. In DCS it's mid 60s to low 70s. This is with ambients of high 20s.

[Hiob]

Just now, edmuss said:

It peaks at about 1030mv I think, I started at 825mv but it wasn't stable much above 1600mhz. I've run a few hours of DCS at +1000 memory without a problem, more testing required at higher though.

Temps wise during heaven it's peaking at mid 70s core with memory a couple of degrees hotter. In DCS it's mid 60s to low 70s. This is with ambients of high 20s.

Ok, when Energy saving is your main goal, a low voltage is fine of course. But you could probably gain quite a lot performance with a voltage in the mid 900s.

Every card is different. Maybe mine is exceptional weak for memory overclocking. On the other hand, the Ti‘s are especially strong/binned for memory performance.

The results may vary for every card, but the approach is always the same. Always change only one variable and test, rinse and repeat. Personally I would start with the GPU and when satisfied continue with Vram.