[SOLVED] Can somebody please tell me what exactly have I done here? Undervolting question

[Noobpunk]

Hey all!

This is my first post so my apologies for the newbie question!

I basically want to know whether I have done the right thing or not..

So my intention is to undervolt my 3090 so I can reduce the temperatures of my GPU; especially the memory junction temperature with the intention of not overclocking anything.

On GPUZ it states
Default Clock - 1395 MHz
Boost - 1860 MHz

What I did next was I loaded up MSI Afterburner and entered -200 on the core clock and created the fan curve of 813mV 1786MHz. Here is the fan curve; https://ibb.co/gtkjRf4

I have been able to reduce temperatures and performance for now seems stable. However now looking at GPUZ it states:

GPU 1198MHz
Boost - 1663 MHz

Although when I am in game on Warzone it states my GPU Clock is 1,800 MHz.

Is everything runnin fine? Could somebody please explain why it appears this way?

Regards,

 

[tennis2]

There's a button on Afterburner (or Ctrl + F) that allows you to see the frequency voltage curve adjustments you've entered. That will visually show you the "stock" curve and the "new" curve.

You want to apply a positive clock offset to effectively apply an undervolt (run xxxx + 200MHz at x.xxx V)

I would also implore you to play around with the % power limit setting. I have my RTX3060Ti running at 70% power limit and reach nearly stock frequency with my +130MHz core clock offset. Roughly a 40-60W power savings

 

[Karadjgne]

Thing called GpuBoost3. It's built into the nvidia drivers. Basically what it does is totally ignore whatever the manufacturer says is any clock set above base, and takes your clocks and voltages up as high as they'll allow, determined by temps and power limits. So you set clocks at 1663MHz, it boosts to 1800MHz because it has the room.

Best thing you can do for a 3090 to lower temps is not mess with clocks or voltages as such, but lower the power limit to 90-95%. That lowers the limits on what GB 3 has to work with, which lowers voltage use, lowers boost clocks and effectively lowers temps with minimal hit to maximum fps.

 

[Noobpunk]

There's a button on Afterburner (or Ctrl + F) that allows you to see the frequency voltage curve adjustments you've entered. That will visually show you the "stock" curve and the "new" curve.

You want to apply a positive clock offset to effectively apply an undervolt (run xxxx + 200MHz at x.xxx V)

I would also implore you to play around with the % power limit setting. I have my RTX3060Ti running at 70% power limit and reach nearly stock frequency with my +130MHz core clock offset. Roughly a 40-60W power savings

In simple terms can you tell me whether I have undervolted safely and not overclocked?

https://ibb.co/gtkjRf4 - Here is the link to the graph!

 

[tennis2]

In simple terms can you tell me whether I have undervolted safely and not overclocked?

You've overvolted by applying a negative MHz offset (thought you can only apply positive MHz offsets, but I'm not at my machine rn). That's why you're now seeing lower frequencies. <MHz at the same voltage (actually it's >voltage at the same MHz compared to stock) means the card is using more power at a given MHz. Given GPU Boost 3.0, you're almost always going to be running at the max set power limit (GPUz has a sensor readout for this)

Simple terms, hit Ctrl + F and look at the frequency-voltage curve as I outlined previously. The stock curve is grey, the modified curve is white.

 

[Noobpunk]

Thing called GpuBoost3. It's built into the nvidia drivers. Basically what it does is totally ignore whatever the manufacturer says is any clock set above base, and takes your clocks and voltages up as high as they'll allow, determined by temps and power limits. So you set clocks at 1663MHz, it boosts to 1800MHz because it has the room.

Best thing you can do for a 3090 to lower temps is not mess with clocks or voltages as such, but lower the power limit to 90-95%. That lowers the limits on what GB 3 has to work with, which lowers voltage use, lowers boost clocks and effectively lowers temps with minimal hit to maximum fps.

Sorry I just got your message now! I really dont mind losing a little bit of performance doing this method just for the temps to drop the amount i have!

 

[tennis2]

Sorry I just got your message now! I really dont mind losing a little bit of performance doing this method just for the temps to drop the amount i have!

Applying an undervolt will just result in a > boost clock because of GPU Boost 3.0 (there may be a limit to how high the frequency will go though). Where you start saving power & heat is when you apply a power limit % setting < 100%

IIRC, my 3060Ti has a stock power limit of 240W, so setting a 70% power limit results in 170W max power draw. The frequency could be slightly lower (depending on the workload), but not 30% less.

 

[Noobpunk]

Applying an undervolt will just result in a > boost clock because of GPU Boost 3.0 (there may be a limit to how high the frequency will go though). Where you start saving power & heat is when you apply a power limit % setting < 100%

Hmm ok, because the tutorial on youtube have specifically all entered -200 on the Core clock option on MSI which then the curvers drop lower than the stockg curve.

Sorry I am a little confused now..

 

[tennis2]

If the new curve is lower than the stock curve (Frequency on the Y axis and Voltage on the X axis IIRC), then trace up from a given voltage and you'll see that the new curve is hitting a lower frequency at 1.0V (for example) compared to stock.
Same applies by tracing a given frequency: 2000MHz is now > voltage on the new curve compared to stock.



The curve below is a properly applied undervolt, see higher frequency at a given voltage OR lower voltage at a given frequency.

 

[Noobpunk]

You've overvolted by applying a negative MHz offset (thought you can only apply positive MHz offsets, but I'm not at my machine rn). That's why you're now seeing lower frequencies. <MHz at the same voltage (actually it's >voltage at the same MHz compared to stock) means the card is using more power at a given MHz. Given GPU Boost 3.0, you're almost always going to be running at the max set power limit (GPUz has a sensor readout for this)

Simple terms, hit Ctrl + F and look at the frequency-voltage curve as I outlined previously. The stock curve is grey, the modified curve is white.

Sorry but I did essentially what was in this video;

View: https://www.youtube.com/watch?v=FqpfYTi43TE


In all the tutorials I have come across, the curve is always below the stock curve, I dont understand why yours is higher?

 

[hotaru.hino]

In all the tutorials I have come across, the curve is always below the stock curve, I dont understand why yours is higher?

Voltage impacts power dissipation exponentially (V ^ 2), whereas frequency impacts it linearly. The GPU also will adjust its frequency based on how much load it has. So if the GPU settles at something like 1,500 MHz, having the point below the stock curve will make the GPU use more voltage to get to that frequency, whereas having it above the stock curve will make the GPU use less voltage. Setting the points below the curve will actually cause the GPU to use more power for a given frequency.

Though this is assuming the GPU will happily go to the voltage in the V-F curve.

 

[Karadjgne]

In English lol. If you run 1800MHz at 1.0v as a default stock setting, you use 100% voltage to reach 100% clocks. If you lower the clock speeds 200Hz, you effectively over-volt, because now you see 100% voltage for 90% clocks. If instead you lower just voltage 0.1v, you now effectively overclock, use 90% voltage to reach 100% clocks. Your clocks at 0.9v are faster than the stock settings which would normally be a 1:1 ratio. Lowering voltage doesn't touch amperage, that remains the same and volts x amps = watts.

By leaving voltage and clocks at default ratio, and lowering the power limit instead, you'll still get some boost above what's set, but since the card is now limited on how much wattage it has to use, that also affects clock boost levels.

 

[tennis2]

yeah, that video is not correct.

See below with reference to 1700MHz. Stock voltage was ~833mV, new voltage is 800mV. THAT's an undervolt. Less voltage at a given frequency.

 

[hotaru.hino]

Lowering voltage doesn't touch amperage, that remains the same and volts x amps = watts.

Nitpicking, power dissipation of a switching transistor isn't measured with ohm's law. It's P = C * V^2 * f. See: https://electronics.stackexchange.com/questions/33318/power-consumed-by-a-cpu (it also has the math to explain it)

This is why undervolting has a much higher impact to power consumption than adjusting frequency.

 

[Phaaze88]

Yikes.
@Noobpunk Try this out, maybe you'll get more out of it. It's how I've undervolted my 1080Ti.

Leave Afterburner + monitoring window running in the background. Don't apply any profiles on the card - let it run at its stock.
Play game/run benchmark. The target: finding max core clock and voltage. Write 'em down or memorize.
Close game/benchmark. Go back to Afterburner and un-link the Power and Temperature Limits. Should be a chain link icon you click on.
Drag Power Limit slider all the way up and then open the Curve Editor. [Raising the Power Limit makes it harder for the core to run into it along with the undervolt.]
That max gpu voltage? Subtract 0.05v from it, find and click the new value/point on the curve. [I simply took an EZ cpu offset value and applied it to this. Trying to find the lowest possible for these gpus is not worth the hassle.]
Raise the frequency on that point to the max core clock, but add 10mhz. [I noticed on a warmer day than usual, the few higher degrees C caused my gpu's 1949mhz to drop to 1936mhz and pretty much stay there until closed. If I raise it a little over, it'll end up dropping to the original max boost.]
Lock it in with the L key, then click the Apply button on the main HUD. Recommend you save the profile for quick and easy application.
If it crashes during testing, instead of -0.05v from max voltage, do -0.04v, -0.03v, and so on.

 

[tennis2]

Lower voltage = lower power draw.
Lower frequency = lower voltage = lower power draw

 

[Noobpunk]

In English lol. If you run 1800MHz at 1.0v as a default stock setting, you use 100% voltage to reach 100% clocks. If you lower the clock speeds 200Hz, you effectively over-volt, because now you see 100% voltage for 90% clocks. If instead you lower just voltage 0.1v, you now effectively overclock, use 90% voltage to reach 100% clocks. Your clocks at 0.9v are faster than the stock settings which would normally be a 1:1 ratio. Lowering voltage doesn't touch amperage, that remains the same and volts x amps = watts.

By leaving voltage and clocks at default ratio, and lowering the power limit instead, you'll still get some boost above what's set, but since the card is now limited on how much wattage it has to use, that also affects clock boost levels.

So what should I do to not overclock but just reduce the amount of power consumption?

Because running Heaven benchmark right now with the settings I have shown everyone, the GPU power draw and GPU Core Voltage is less than the stock settings, thus leading to lower temps. My apologies for this mess, I would rather learn this now and later. Am I damaging my GPU at all by doing this?

So basically everyone on youtube following these videos are being misguided?

Running Heaven Benchmark;

Stock Core Clock: 1,900 MHz
Stock Core Voltage: 1,056V
Stock GPU Power: 327 W

When I enter -200 on Core Clock and adjust the fan curve to 813mV 1786MHz

GPU Clock: 1,800 MHz
GPU Core Voltage: 0.813 V
GPU Power: 260-80 W

As well as all of my GPU temperatures reducing by a bit too.

 

[hotaru.hino]

So what should I do to not overclock but just reduce the amount of power consumption?

Because running Heaven benchmark right now with the settings I have shown everyone, the GPU power draw and GPU Core Voltage is less than the stock settings, thus leading to lower temps. My apologies for this mess, I would rather learn this now and later. Am I damaging my GPU at all by doing this?

So basically everyone on youtube following these videos are being misguided?

Running Heaven Benchmark;

Stock Core Clock: 1,900 MHz
Stock Core Voltage: 1,056V
Stock GPU Power: 327 W

When I enter -200 on Core Clock and adjust the fan curve to 813mV 1786MHz

GPU Clock: 1,800 MHz
GPU Core Voltage: 0.813 V
GPU Power: 260-80 W

As well as all of my GPU temperatures reducing by a bit too.

What's the goal here? If reducing power consumption is the goal, you can do that by setting the power limiter. You don't have to undervolt or anything. If you're trying to maximize the efficiency of your GPU, then yes, you have to undervolt it by adjusting the V-F curve so the overall curve is above the stock curve.

 

[tennis2]

So basically everyone on youtube following these videos are being misguided?

Don't blindly believe everything you see on the internet. I've already explained how to verify that you've offset the frequency voltage curve properly for an undervolt. Notice in the video you linked (at 2:30 which is what I think you're talking about), he actually drags the singular point ABOVE the stock curve line before hitting apply (and hence capping the frequency as a horizontal line). Therefore, he effectively undervolted by 100mV for the max frequency AND set a frequency limit of 1850MHz but in the process of doing so, overvolted almost every frequency point below that. He hit apply pretty quickly, so you'll have to pause-step to catch it. The singular point ended up at about +150MHz offset from the stock curve, which is about the maximum stable value for most RTX30xx GPUs from my research.

The reason he had to drag the curve down at the start was to get all the points to the right of his frequency cap to be BELOW that point. That's a fault of how MSI afterburner works to extrapolate that horizontal line by only dragging a single point. It only does this if the singular point is above the remaining curve.

The reason to apply a frequency cap is to limit frequency deviation during gaming. Because of GPU Boost 3.0 and how different work loads draw different amounts of power (ie game compared to OCCT burn test, or just different parts of a gameplay) you can get pretty wide frequency swings if you're running <<<100% power limit. Capping the frequency reduces that deviation. Certainly more of an advanced (and painstaking having to drag all the points properly in Afterburner to maintain your undervolt below the frequency cap) methodology that should/can be avoided unless you find frame rate variance is being bothersome. I've got a 60% Power Limit profile that I applied a frequency cap to since clock speeds could vary between 1600MHz to 2100MHz in-game (30% swing) if frequency was uncapped.

So what should I do to not overclock but just reduce the amount of power consumption?

<100% power limit = reduce total power draw

 

[Noobpunk]

Don't blindly believe everything you see on the internet. I've already explained how to verify that you've offset the frequency voltage curve properly for an undervolt. Notice in the video you linked (at 2:30 which is what I think you're talking about), he actually drags the singular point ABOVE the stock curve line before hitting apply (and hence capping the frequency as a horizontal line). Therefore, he effectively undervolted by 100mV for the max frequency AND set a frequency limit of 1850MHz but in the process of doing so, overvolted almost every frequency point below that. He hit apply pretty quickly, so you'll have to pause-step to catch it. The singular point ended up at about +150MHz offset from the stock curve, which is about the maximum stable value for most RTX30xx GPUs from my research.

The reason he had to drag the curve down at the start was to get all the points to the right of his frequency cap to be BELOW that point. That's a fault of how MSI afterburner works to extrapolate that horizontal line by only dragging a single point. It only does this if the singular point is above the remaining curve.

The reason to apply a frequency cap is to limit frequency deviation during gaming. Because of GPU Boost 3.0 and how different work loads draw different amounts of power (ie game compared to OCCT burn test, or just different parts of a gameplay) you can get pretty wide frequency swings if you're running <<<100% power limit. Capping the frequency reduces that deviation. Certainly more of an advanced (and painstaking having to drag all the points properly in Afterburner to maintain your undervolt below the frequency cap) methodology that should/can be avoided unless you find frame rate variance is being bothersome. I've got a 60% Power Limit profile that I applied a frequency cap to since clock speeds could vary between 1600MHz to 2100MHz in-game (30% swing) if frequency was uncapped.


<100% power limit = reduce total power draw

I appreciate the response. I will reduce the total power draw and see what temps I get as well. But I am not looking to over clock at all.

but please answer this, am I damaging my gpu in anyway in doing what I did?

 

[hotaru.hino]

I appreciate the response. I will reduce the total power draw and see what temps I get as well. But I am not looking to over clock at all.

but please answer this, am I damaging my gpu in anyway in doing what I did?

The only way you can damage your GPU is if it exceeded any electrical limits. And even if you told Afterburner to say exceed 1.5V or over 100% power, these are really suggestions and the GPU's firmware has soft limits in place to prevent stuff like this from happening.

 

[Noobpunk]

The only way you can damage your GPU is if it exceeded any electrical limits. And even if you told Afterburner to say exceed 1.5V or over 100% power, these are really suggestions and the GPU's firmware has soft limits in place to prevent stuff like this from happening.

Ok so in my case the gpu will not be damaged in doing what I did?
My apologies again..

 

[hotaru.hino]

Ok so in my case the gpu will not be damaged in doing what I did?
My apologies again..

While I won't make a guarantee, everything you described what you did was within the limits of the video card, so there was almost no chance of something bad happening.

 

[Noobpunk]

While I won't make a guarantee, everything you described what you did was within the limits of the video card, so there was almost no chance of something bad happening.

and if I continue in doing so in this matter, I should be ok? Because I am not over clocking nor am I doing anything out of its limits?

 

[USAFRet]

So my intention is to undervolt my 3090 so I can reduce the temperatures of my GPU

Maybe I missed it, but I don't see any temperature number associated.