[SOLVED] 3080FE Undervolting giving less performance?

[ps2cho]

Hi all,

I just finished with my RAM timing/overclocks so now moved to my GPU. I have not undervolted before (ITX case), however I've ran into a weird issue.

Why is my undervolt decreasing performance? If you look at the voltage/frequency chart I edited, I am asking the GPU for the SAME clock speeds at lower voltages. I dont understand why that would decrease performance? Shouldn't it keep everything the same?

Stock 3080 FE according to Afterburner is 1130mv Core @ 2,025mhz core clocks.
I undervolted to 956mv @ 2025mhz core, and temps didnt move at all (both 72.8-73C max), and the score went down.

So I am confused. What am I missing here?

 

[helper800]

I am not very experienced on this matter but it stands to reason that the gpu cannot achieve those frequencies with the given voltages. Have you seen a 1:1 increase in score with an increase in voltage? For example try 1050mv at 2025 as a test?

 

[ps2cho]

I am not very experienced on this matter but it stands to reason that the gpu cannot achieve those frequencies with the given voltages. Have you seen a 1:1 increase in score with an increase in voltage? For example try 1050mv at 2025 as a test?

yes, correct. The more volts I give, the higher the score until it returns back to stock. Even a small 50mv decrease drops the score in a direct correlation. Its as if the GPU wont tolerate ANY undervolting, which doesnt seem right, so I figure I'm doing something wrong here.

For instance here is 1,075mv @ stock 2,025mhz core. You can see its still less than the stock high score of 30,910....so somethings wrong!

EDIT:

Actually when I look in GPUz, it is showing difference frequencies. Is the voltage curve in MSI Afterburner maybe wrong?

 

[dotas1]

The undervolting I did to my card taught me some things. One of them is that when you ask the GPU to run at a specific clock at a specific voltage, it doesn't mean that it will achieve that clock. It will try and reach as high as it can for that voltage.

For example I tried to reach 2000Mhz at 890mV. It did not reach and settled at 1900ish. When I tried 850mV and 1950Mhz, I had crashes , so it could not achieve clocks within the acceptable margin.

Keep in mind that lowering the voltage and asking lower clock, will result in lower temperatures which MAY result in higher clocks due to thermal headroom (that's how GPU boost works actually).

 

[helper800]

yes, correct. The more volts I give, the higher the score until it returns back to stock. Even a small 50mv decrease drops the score in a direct correlation. Its as if the GPU wont tolerate ANY undervolting, which doesnt seem right, so I figure I'm doing something wrong here.

For instance here is 1,075mv @ stock 2,025mhz core. You can see its still less than the stock high score of 30,910....so somethings wrong!

EDIT:

Actually when I look in GPUz, it is showing difference frequencies. Is the voltage curve in MSI Afterburner maybe wrong?

In my opinion there is nothing wrong. Undervolting is not magical switch. Typically undervolting only helps performance when GPU temperatures are higher. This is a direct result of how boosting works in relation to temperature. The lower the temps the higher clock speeds can go given adequate voltage. Since your card is not running too hot decreasing voltage decreases achievable clock rates. The only thing undervolting should do for your case is decrease temperature, but since you are also trying to achieve similar clock speeds temps stay the same. As a general rule of thumb with GPUs; Clock speeds are limited by power, both wattage limits and voltage regulation, and temperature. As temperature decreases boosting try's to get as much clock speed out of its power and vRel limits as possible. If the GPU is cool enough to boost to the highest clock but you reduce the voltage the clock speeds, as a result, will also be reduced. Now, like I said, I am the most adequately experienced in this area, I could be missing something.

 

[faalin]

If you look at both GPU Clock Frequency charts you can see the one where you limited the voltage there are a lot more pronounced dips in it. The 2000/3000 will boost themselves if the temps and voltage is there. By limiting the voltage the card cant reach the same Mhz all the time as the stock settings. As long as the card is still staying within temps just leave the voltage alone, If your not trying to inch every single Mhz out of the card just let it manage its self.

 

[ps2cho]

If you look at both GPU Clock Frequency charts you can see the one where you limited the voltage there are a lot more pronounced dips in it. The 2000/3000 will boost themselves if the temps and voltage is there. By limiting the voltage the card cant reach the same Mhz all the time as the stock settings. As long as the card is still staying within temps just leave the voltage alone, If your not trying to inch every single Mhz out of the card just let it manage its self.

ok I guess that makes sense - what I was trying to do was see if I could maintain the same performance, with less heat generated. So this isn't really possible because of the boost algorithms?

 

[jasonf2]

Undervolting has more to do with thermal control than working performance. So if you have a card that has a thermal solution that is doing its job well there is a very good chance that reducing the voltage will actually negatively impact frequency. When overclocking, a voltage increase, not a decrease is usually needed to boost clock. The underclock has became a thing lately only because many of the nvidia cards run stock hot and have a habit of thermal throttling (especially in laptops) and that mining can sometimes get a efficiency sweet spot going by tweaking just a bit. But other than that unless you have a thermal issue an undervolt will not help performance.

 

[helper800]

ok I guess that makes sense - what I was trying to do was see if I could maintain the same performance, with less heat generated. So this isn't really possible because of the boost algorithms?

This can be possible, however, you run into power bottle necks before you run into heat dissipation issues. The amount of performance gained from boosting at a lower temperature does not compensate you for the reduced vRel. Your particular GPU is power starved before it gets too hot.

 

[ps2cho]

ok thanks - I was just most interested in what others have done:
https://bjorn3d.com/2020/10/undervolting-the-rtx-3080-and-the-rtx3090/2/#split_content

See how in their undervolt it dropped wattage almost 80w of power with minimal impact in FPS. 74C to 68C max load. I tried copying their exact inputs and I still get a significant decline in performance.

 

[helper800]

ok thanks - I was just most interested in what others have done:
https://bjorn3d.com/2020/10/undervolting-the-rtx-3080-and-the-rtx3090/2/#split_content

See how in their undervolt it dropped wattage almost 80w of power with minimal impact in FPS. 74C to 68C max load. I tried copying their exact inputs and I still get a significant decline in performance.

If you want to experiment a bit put the cards settings back to stock and adjust the power limit to about 85-92%. You will notice that the temps are less and clocks will drop by about 50mghz but you will be running a few degrees less on the core. Watts, vRel, and temps all have direct impacts on performance. None of examples in the link provided were able to keep performance the same as before. Generally adjusting the power limit of the card (Watts) will effect temperatures the most. If you want a quite system then reducing the power or vRel is all you can do at the same fan speeds to achieve the goal.

 

[Phaaze88]

Wow. That undervolt is huge. 1130 to 956mv = 174mv total?
Then again, I'm looking at it from Pascal(1080Ti) and not Ampere. I didn't even get over 100mv before I started seeing stability issues.
It also doesn't help if you're only looking at this through 3D Mark. It is not a one size fits all application. You have to test as many games/benches as possible. Other games may crash or cause voltage throttling(clocks drop due to not enough voltage).

In the end, I settled for a 50mv undervolt. It's been pretty smooth so far, besides, that's a simple one that seems to work for Intel cpus too.

 

[Karadjgne]

Throw out the base 2025MHz base frequency. It's just that, a base. Modern nvidia gpus use GPU Boost 3 which can and will boost far above the base frequency depending on 2 things. Temp and power. If you lower power in the card, it'll lower the ability to boost. If you lower temps, it'll raise the ability to boost. Upto a stable limit.

So even when your base remained the same, the large undervolt starved the card, so it didn't boost as high, which lowered your score. Leaving the power where it is, if you could lower temps on the card, it might boost back to where you started, might not.

. I tried copying their exact inputs and I still get a significant decline in performance.

That's because your card is different. Doesn't matter if it's the exact same brand and model, the silicon is different, voltages and current and resistances inside the card are slightly different, tolerances are slightly different, so using exact same settings will invariably end up with different results.

 

[ps2cho]

I did just try dropping the power limit to 90% and it also decreased my 3dmark score, so it does appear there basically is no option on my card to reduce power consumption or heat without sacrificing performance. Guess that's one less thing to worry about then!

 

[Karadjgne]

Define performance. 3DMark doesn't do it. 3DMark defines Limits, not the journey to get there.

Your gpu is going to mostly be well under 100% usage, and rarely if ever hit the same temps as 3DMark punishment shows. So you list 1-3fps at the maximum. That's not really going to affect the in-betweens.

I could OC my 3770k to 5.0GHz at 1.42v. Dropped it back to 4.9GHz at 1.32v and dropped max temp as a result. That didn't change my idle temps, didn't change gaming temps and I honestly couldn't see any difference.

With such a small degree of score change, just outside margin of error, you'd not see any difference either. It's not always the best course of action to drag every erg of performance out of a component. Sometimes it's a balance of what works vs what you can live with.

 

[jasonf2]

I did just try dropping the power limit to 90% and it also decreased my 3dmark score, so it does appear there basically is no option on my card to reduce power consumption or heat without sacrificing performance. Guess that's one less thing to worry about then!

That really isn't a bad thing. It just means that your card manufacturers engineers did their job well! If you were able to pull off a 10% efficiency gain for no performance loss the engineers would have been leaving way too much on the table. You have to remember that the manufacturer's engineers set things up the way they do for a reason. And no matter what the whole thing is a balancing act between performance, efficiency and stability.

 

[Pantreus]

Hi all,

I just finished with my RAM timing/overclocks so now moved to my GPU. I have not undervolted before (ITX case), however I've ran into a weird issue.

Why is my undervolt decreasing performance? If you look at the voltage/frequency chart I edited, I am asking the GPU for the SAME clock speeds at lower voltages. I dont understand why that would decrease performance? Shouldn't it keep everything the same?

Stock 3080 FE according to Afterburner is 1130mv Core @ 2,025mhz core clocks.
I undervolted to 956mv @ 2025mhz core, and temps didnt move at all (both 72.8-73C max), and the score went down.

So I am confused. What am I missing here?

I haven't read the responses so this may have already been covered, but there are some issues with the undervolt. First, while you have undervolted your target frequency, which is what you'll hit under heavy load, you have effectively overvolted your entire curve to the left of that node. The method you have used is what we see all over YouTube, but by starting with a negative offset to lower the curve, you are achieving the exact opposite of what you were aiming for. For example: If your stock curve had 1700MHz being targeted at .900mV, then by dragging that entire curve down by 300MHz, you are now targeting 1400MHz at .900mV. So, when you are not under maximum load, your GPU is actually much less power efficient.

Now, I think the reason you are seeing lower performance in benchmarks is because your effective clock is considerably lower than your core target clock. If you use an on-screen display, such as Afterburner/RivaTuner, GeForce Experience etc.. the clock you are seeing is your target clock and not what your GPU is actually operating at. If you check HWiNFO, there is an effective clock sensor. Watching that while performing a benchmark will always show that your effective clock is lower than your core clock. Ideally, your effective clock won't be more than about 10MHz below your core clock. Adjusting the voltage curve as you have is not optimal in this regard.

What I would recommend, at the very least, is to start with your stock curve, shift-click on the 956mv node and drag the entire curve up until that node has reached 2025mhz. Then, you can shift-click in an open space to the right of that node and drag your cursor to the right, highlighting all of the nodes above it, then drag them all down and click "apply" to flatten the right side of your curve.
Possibly even better is (this is what someone else recommended I do) to first run the OC Scanner and save it to a profile. Then use that curve as your starting point, rather than the stock curve. The OC Scanner curve will be kind of janky looking and if you shift-click to drag a node up, it'll reset the shape. You'll instead need to alt-click a node to drag the curve up so the shape is retained.

I can't speak to your GPU or your setup, but another issue is likely thermals. Your GPU isn't actually going to boost all the way to 2025MHz unless your temps are super low. My 3080 runs hot on stock settings and it's impossible to actually hit the target peak clock no matter how I handle the voltage. In your case, with that lowered curve, you're asking your GPU to target lower frequencies than normal when it's unable to hit the highest target frequency.

Here's a run at stock settings and also with an undervolt and slight underclock:

And here's the profile I use most often, which is a more significant undervolt and underclock. However, it performs almost as well as my performance profile but at significantly lower temps and power draw.

TL;DR: When adjusting your curve, don't start with a negative core clock offset and then spike it at the voltage you want it to maintain. Either start with your stock curve or a curve determined by the OC Scanner, pick a voltage you want and then drag the entire curve up until that node meets the frequency you want to target, then just flatten the curve to the right of that node.

P.S. I know that my CPU makes absolutely no sense when I'm running a 3080.

P.P.S. Not exactly on topic, but I've done some fairly thorough testing with VRAM overclocking in Superposition and can say there is almost zero benefit to be had from overclocking VRAM, at least from a gaming perspective.

 

[Karadjgne]

All gpus have a power target. That's what the gpu wants in order to output whatever clocks it uses. Power is volts x amps. If the card has a higher power target than required for the clocks, lowering voltage can lower temps and allow higher boosts for longer, raised performance but still meets power targets for the clocks.

To a point. Lower voltages too much and one of 2 things happens. Either amperage goes up to meet the demand, which raises temps in VRM's , or boost comes down to match power levels and subsequent loss of performance, even at lower temps.

It's a fine line to walk and results are not always visible as many gpus do not have temp sensors on other components or any way to realistically read those sensors outside of the card if it does, only the gpu itself has a sensor on die. So 2 gpus can read 70°, but one has VRM's at 90°, the other has VRM's at 99°, same performance ±.

 

[Phaaze88]

^Yeah, some users are just taking the voltage too low for the desired core clock.
Vram OC'ing is not too forgiving either - it's best left alone.


Aww crap, this is a thread from 4 months ago...