Overclockers: Why You Should Undervolt Your Graphics Card.

[edmuss]

I did try quickly at 950mv but it seemed to crash heaven without too much hassle. I'll give it a bit more of a thorough test to see if I can eke some more core out of it!

The asus tuf cards do have a separate VRAM heatsink so I guess that might help decouple the thermals from the core, not too sure how much difference it would actually make or is it just asus marketing?

I've not played with gddr6x before but when overclocking the gddr6 on the 3070 it would produce artifacts on anything much above +1200mhz. I know that it has ecc functionality and will error correct in the background whilst slowing down slightly, also it seems to have a false ceiling that will produce more errors and slow down before then speeding up.

It could be that at +1000mhz (which gives ~205fps at the start of heaven) is below the ceiling and at +1500mhz (which gives ~220fps at the start of heaven) is above the ceiling.

Edited September 9, 2022 by edmuss

[Hiob]

1 minute ago, edmuss said:

I did try quickly at 950mv but it seemed to crash heaven without too much hassle. I'll give it a bit more of a thorough test to see if I can eke some more core out of it!

The asus tuf cards do have a separate VRAM heatsink so I guess that might help decouple the thermals from the core, not too sure how much difference it would actually make our is it just asus marketing?

I've not played with gddr6x before but when overclocking the gddr6 on the 3070 it would produce artifacts on anything much above +1200mhz. I know that it has ecc functionality and will error correct in the background whilst slowing down slightly, also it seems to have a false ceiling that will produce more errors and slow down before then speeding up.

It could be that at +1000mhz (which gives ~205fps at the start of heaven) is below the ceiling and at +1500mhz (which gives ~220fps at the start of heaven) is above the ceiling.

As I said, when testing for the optimal clock speed, I would leave the Vram at stock. 100 MHz core clock will give you more fps gain than 1000 MHz Vram. (at least that’s the case for non-ti‘s)

[edmuss]

Yeah I'll be winding the VRAM back for higher core testing, it's a quick and easy one to change compared to adjusting the curve if afterburner is deciding to be flaky

[EightyDuce]

It's going to depend card to card and as been mentioned below 900 mV is pretty aggressive under-volting so I instability is not unexpected. As for memory, it is also a silicon lottery as I'm easily able to hit +1200 on the vRAM and no crashes in DCS; max stable your core first then do RAM as core will bring the most performance uplift vs. the vRAM. At least on a 3060ti.

[edmuss]

Well I've spent the last 90 minutes watching runs of heaven benchmark

Result is, 1845mhz @875mv +1000 memory is best bang for buck and at least 5°C cooler than any other settings.

I tried up to 1980mhz @975mv (in 15/25 increments) and it just didn't hold the core clocks but was getting up to 86°C.

I've already proved this setting DCS stable for a few hours so I guess I just carry on and see what happens

[Hiob]

8 hours ago, edmuss said:

Well I've spent the last 90 minutes watching runs of heaven benchmark

Result is, 1845mhz @875mv +1000 memory is best bang for buck and at least 5°C cooler than any other settings.

I tried up to 1980mhz @975mv (in 15/25 increments) and it just didn't hold the core clocks but was getting up to 86°C.

I've already proved this setting DCS stable for a few hours so I guess I just carry on and see what happens

That‘s perfectly fine. The ultimate goal is to find the settings that suits you best. From the temperatures you mentioned, I see that your card is air cooled. In this case undervolting without raising clocks may very well be the better approach. 

I wrote somewhere else earlier, that the amount of thermal headroom (water, air, case air flow, ambient temps) is very important for the kind of results you can expect. Reducing the Voltage, and therefore the energy consumption and temperatures is always the best starting point and works on air and water. What you‘re going to do with the gained headroom is a matter of taste (and cooling and silicone lottery). Either keep it economic or try and see how much you can exchange back for higher clocks.

I would suspect, with such low voltage and high Vram clocks, that you may encounter problems with other games that put a very different load on the GPU. (Try Cyberpunk). But that’s just speculation. Maybe I‘m mistaken.

[edmuss]

I am indeed air cooled, also in a small enclosed space so the ambient ramps up to almost 30 pretty quickly once the GPU starts acting as a space heater

Good shout with cyberpunk, it's installed so no issue to go test! I am only on 1080p though so I suspect that DCS vr will be so much more taxing.

[Lange_666]

For me it's not about OC the card but more on generating less heat (and as a surplus use less energy), the 3080Ti runs quite hot even at default (60°C).
I'm currently down to 975mv, 950mV give problems in VR.
Another thing is noticed is that i can't run the Afterburner Scan test without the card crashing, not even on default values.

Did see another undervolting video last night where the guy always talked about megaVolt instead of milliVolt... LOL, hew knew that it was MegaHz instead of milliHz though.  

Edited September 10, 2022 by Lange_666

[MAXsenna]

22 minutes ago, Lange_666 said:

Did see another undervolting video last night where the guy always talked about megaVolt instead of milliVolt... LOL, hew knew that it was MegaHz instead of milliHz though.  

[edmuss]

43 minutes ago, Lange_666 said:

For me it's not about OC the card but more on generating less heat (and as a surplus use less energy), the 3080Ti runs quite hot even at default (60°C).
I'm currently down to 975mv, 950mV give problems in VR.

Same here but at the same time maximise the performance.

Are you saying that at the card is idling at 60°C? That is incredibly hot, even with my fans stopped it's sitting below 40. Do you have a particularly small case or poor airflow? If your idle temps are that high then load temps will be through the roof which will likely be a cause for instability.

[BitMaster]

60°C in idle ?   There is a problem

[Lange_666]

Yes, idling at around 60°. Brand is EVGA and looking through the internet i found the same story all over the place for their 3080Ti's.
Fans don't run before it hits 60°.
I set a manual curve through Afterburner which makes the fans spin a bit at idle and that brings the temps right down to below 40° on idle.
In DCS with stettings for 80% maxed out and my manual curves, card hits around 65-70° now and then the fans don't run max.

For the moment with the fans spinning a bit, idle sits around 37-38° with still a higher then normal ambient room temp (the heat still sits in the house).

Case is a Cooler Master big tower but it's a around 7 years old running a i7 4770K. Problem with the case is that it will not fit cards bigger then 290mm in length and the EVGA 3080Ti XC3 was the only one that would fit since it's only 285mm. I would have gone for a 3090 for the 24GB VRAM but none of these are shorter then 3xxmm so i ended up with the EVGA.
Overall idea was to buy an entirely new puter next year but recent DCS and MSFS updates showed a slowly degradation of performance and a GPU that always maxed out in MP so i decided to jump a bit ahead and update the graphics card now and then use that on a later build early next year.
However, as of yesterday i'm looking to do the upgrade sooner because i had some temp issues on my i7 4770K the last days coming from the watercooler. There was some dirt clogging up a good waterflow (like here: https://youtu.be/BJ2ELhS6kts?t=108 ). Fixed now but still not 100% good since i can't run the same overclock on the CPU.
Still have to figure out what to do, either wait for the Ryzen 7000 and AM5 / Intel 12 or 13 series and boards with DDR5 (more future proof) or go AMD 5800X3D with an AM4 DDR4 board (which i guess will be cheaper after the 7000 series is released).

 

Edited September 10, 2022 by Lange_666

[Hiob]

Water cooling does make a lot more sense for GPUs than for CPUs. CPUs in general have a pretty poor heat conductivity in opposite to GPUs due to the heatspreader and overall design.

Custom watercooling is of course a challenge and cost factor on its own, but very rewarding for GPUs.

Mine is idling at ambient temperature and under load hits 12-15 C over ambient. That’s about 44 C in the worst case, whereas the CPU (same custom water loop) will still run in the 60s and spiking up to 80 C on water. 

[edmuss]

Still seems hot, mines idling at 30/31°C with a 23°C ambient.

Have you got a good air flow through the case?

 

Edited September 10, 2022 by edmuss

[Lange_666]

The airflow is OK, maybe not 100% optimal (i'm no expert) but it worked fine in my previous setup.
Currently i have EVGA's Precision X1 software running and if i set everything to default, fans on the card don't start spinning before the GPU temp hit's 61°.
If i manually set it up that they are spinning around 40% on idle (GPUZ shows more then Precision X1), GPU temp goes down to 33° (ambient 23,5°).
The difference is that your fans don't spin at idle. I might disassemble the card and take a look at the thermals.
What brand do you have?

[img]https://i.imgur.com/9Rf1yj1.png[/img]

After discovering the high idle temps i started to look around if that was normal and found lots of people reporting the same for the EVGA cards.
2 examples:

https://forums.evga.com/3080-XC3-Ultra-60C-idle-temps-m3335501.aspx

https://www.reddit.com/r/EVGA/comments/mdl76s/xc3_ultra_3080_idle_temps/

 

 

 

 

Edited September 10, 2022 by Lange_666

[edmuss]

I've got an asus 3080ti TUF OC, I don't have a case as such so the GPU is open to ambient, I do have a 140mm fan feeding cool air to the GPU though.

[BitMaster]

+1 for WC-GPU.  You rather watercool the GPU these days than a CPU.

The big problem is, finding a matching block for the myriad of AIC designs, that has to be thought of before you click the BUY button.

The hottest I have seen my GPU, tho an old 1080ti, is ~52°C when fully stressed. Idle is a few °C above ambient air...and it's very silent, even under full tilt load.

[Hiob]

1 hour ago, Lange_666 said:

The airflow is OK, maybe not 100% optimal (i'm no expert) but it worked fine in my previous setup.
Currently i have EVGA's Precision X1 software running and if i set everything to default, fans on the card don't start spinning before the GPU temp hit's 61°.
If i manually set it up that they are spinning around 40% on idle (GPUZ shows more then Precision X1), GPU temp goes down to 33° (ambient 23,5°).
The difference is that your fans don't spin at idle. I might disassemble the card and take a look at the thermals.
What brand do you have?

[img]https://i.imgur.com/9Rf1yj1.png[/img]

After discovering the high idle temps i started to look around if that was normal and found lots of people reporting the same for the EVGA cards.
2 examples:

https://forums.evga.com/3080-XC3-Ultra-60C-idle-temps-m3335501.aspx

https://www.reddit.com/r/EVGA/comments/mdl76s/xc3_ultra_3080_idle_temps/

 

 

 

 

 

Seems that they opted for silence over coolness. Which is ok I think. 60C will certainly not hurt your card. With cap on the voltage, the temp may even come down a bit. Personally I would find the fan-% that is still pleasant to my ears and lower the idle temps a bit.

But temps under load are more important!

[EightyDuce]

I would try make a custom fan curve in MSI AB if your fans don't come on until 60C (very strange), which seems rather high, especially for idle.

[Lange_666]

14 hours ago, EightyDuce said:

if your fans don't come on until 60C (very strange)

Looking at their forum it seems normal for 3080Ti cards. Custom fan curve does the job.
 

[Burt]

I noticed that in Nvidia Control Panel in Global settings if you have your Power Management mode set to High Performance it will cause high idle temps 56-60c. If switched to Normal idle will come down to 30ish. I only have High Performance modes set only in Program Settings for DCS / VR .

Also, I’m not a big fan on fan curves. I set my fans on a Manual percent while gaming/DCS. I have one of the most choked out air cases on the market and the fan curves doesn’t do it no good… up - down up down !!! Manual fans settings does why better in keeping good temps and frequencies stable. Just me .02 cents.

Edited September 12, 2022 by Burt

[Lange_666]

My Power Management setting in global settings is always default = Normal so that's certainly not the case of the problem. Or maybe the driver itself is the problem and Normal isn't Normal but Max Power. I'm on the latest 516.94 (clean installed after DDU'ed the one i got installed previously with my 1080Ti card).
I only use Max Performance in my game profiles.

[Hiob]

High performance energy profile in nvcp would be a waste of energy and does exactly nothing for performance. Your just blowing 60W worth of hot air around for nothing.

A custom fan curve (which is pretty easy to do) would certainly do the trick.

[Burt]

6 hours ago, Hiob said:

High performance energy profile in nvcp would be a waste of energy and does exactly nothing for performance. Your just blowing 60W worth of hot air around for nothing.

A custom fan curve (which is pretty easy to do) would certainly do the trick.

Well Im just following what I see every VR user is setting their power management at which is High Performance mode. Why have a curve when at load/gaming your gonna be around 65-75c ? What the curve gonna do …Just manually set your fans to 80-85% same thing to the cpu rad fan. 
 

 

[Hiob]

4 hours ago, Burt said:

Well Im just following what I see every VR user is setting their power management at which is High Performance mode. Why have a curve when at load/gaming your gonna be around 65-75c ? What the curve gonna do …Just manually set your fans to 80-85% same thing to the cpu rad fan. 
 

 

Yeah, the idea is, to keep the card from switching modes during gameplay, which may (or may not) introduce fluctuations in fps. And I certainly don't want to speak for VR as a special case. But the reality is, with the current drivers, once you are in game, the card is under full load and won't reduce power anyway. It is easy to monitor this with the usual tools. It simply doesn't make ANY difference.

But the downside is, instaead of idling between 20-30 Watts on the desktop (values are from my 3080, maybe slightly higher or lower for other cards), it is idling between 80-90 Watts on the desktop. And I don't know about you, but my system certainly spends a lot of time in idle. Do you feel comfortable blowing 60 Watts of energy out for nothing? I certainly don't.