3060ti undervolting

[Taraquin]

Hi, wanted to share my experience with UV my Palit 3060ti dual OC via Afterburner using Superosition benchmark at 1080p extreme and SOTTR@1080p highest with DX12 and TAA. I used the curve editor and powerlimiter. I updated the scores with the max 105% powerlimit and tested a bit more. The stated clockspeed is what I set it to in afterburner, BUT since the card has a stock OC of 30MHz I get real life frequency of 30MHz all the time in all games and Superposition.

Stock
6877 - 206W max - 80C max - fan@50%
-80% powerlimit (voltage stuck around 850mv, frequency around 1665MHz))
6504 - 165W max - 69C max - fan@50%
-1575MHz@700mv (useless as memoryclock gets stuck at 10GHz)
5562 - 119W max - 58C max - fan@30%
-1620@731mv (lowest voltage which allows for 14GHz dram clock)
6041 - 128W max - 65C max - fan@40%
-1665@750mv 105% powerlimit
6349 - 137W max - 67C max - fan@44%
-1800@806mv 105% powerlimit
6748 - 158W max - 65C max - fan@41%
-1905@881mv 105% powerlimit 16GHz dram
7316 - 208W max - 76C max - fan@70%

Anything above 881mv results in powerthrottling and barely gives better scores, but temp and noise increases. Undervolting by curve editor in MSI Afterburner is superior to powerlimit. Even with powerlimit voltage is 100mv above what you can do with curve editor.

In SOTTR results were a bit different:
-Stock@950-1081mv
152fps gpurender - 204W max - 75C max - fan@67%=40dB
-1650@743mv
134fps gpurender - 120W max - 63C max - fan@37%=33dB
-1785@806mv
144fps gpurender - 145W max - 66C max - fan@41%=34dB
-1920@881mv
156fps gpurender - 184W max - 71C max - fan@56%=37dB

For low temp\noise\consumption 1665@750mv gives over 90% of performance for 40% lower consumption in Superposition and is inaudible and seems like the best setting for this as dram clock is kept at max, going down to 725mv dropped memclock by 4000MHz which plummets performance. 1800@850mv gives stock performance at 10% lower consumption, 3C lower temp and a bit less noise.

In SOTTR 1650@743mv gives you 88% of stock performance@ 7dB lower noise and 41% lower consumption. The OC-profile of 1920@881mv gives 3% more performance at 10% lower power consumption.

 

[Arcsaber]

Anybody not overclocking should be undervolting, IMO.

Did you do yours with manual curve adjustment? Because I did it the lazy way: set the power limit slider and use OC Scanner
You can keep dropping and testing until you match the stock performance, but I stopped at 85% - 2-3% performance increase for 15% power decrease.

 

[Taraquin]

Anybody not overclocking should be undervolting, IMO.

Did you do yours with manual curve adjustment? Because I did it the lazy way: set the power limit slider and use OC Scanner
You can keep dropping and testing until you match the stock performance, but I stopped at 85% - 2-3% performance increase for 15% power decrease.

I used curve editor in afterburner, dropping pwr-limit works fine aswell, but you can fine tune it a bit further with curve editor. Pwr-limit is a great choice for getting good results fast and you don`t compromize stability like with curve editor

 

[Arcsaber]

I used the safe way because I could get curve editing stable in Heaven, but Resident Evil 2 would crash in-game.

I might try curve-editing again and test with the OC Scanner's test function.

 

[nasky]

You're missing on some perf though, my Gainward would boost to 2100Mhz but if you're not playing in 4k should be enough I guess

Your score @850mV is basically at 2070S level ...

 

[Chrispy_]

I undervolt everything I can, within reason.

Perceived fan noise does not scale with performance. Something can be completely inaudible at 1500MHz and so obnoxiously loud at 1900MHz that you can't even tolerate being in the same room. There's always a happy medium and for me that's "how fast can this card/CPU go before I notice the fan noise"

My approach to new GPUs is quite simple really;

1. Find out the maximum fan speed I'm prepared to tolerate (usually much lower than stock fan speeds)
2. Find out how slow the card runs at stock voltages at this fan speed, and what power consumption it's using
3. Use OCCT to accurately test the clock/voltage absolute error limit within a range of clocks starting from the clockspeed answer to step 2 and ending at stock voltage, whatever that is.
4. Add 5% for stability and apply this new voltage curve
5. Use OCCT again to monitor power consumption vs clock using the new voltage curve
6. Set defaults clocks that match the power consumption of step 2.
7. Open a beer, pat myself on the back for half an hour of work that nobody except me will appreciate, and fire up <insert Steam game of the moment>

 

[Arcsaber]

I had an old 8350K that ran 4 GHz at 0.98 V. After I delidded it it sat under a passive NH-D15.
Can't wait to get my hands on a 3050 Ti and see where the floor is with that thing...

 

[pavle]

Very good results there, I like to make my graphics cards low_power too.
Too bad one can't edit BIOS of the graphics card as one could with up to and including Maxwell (GTX 980) and just have a low_power card without an additional program running in the background (afterburner or similar).

 

[Taraquin]

You're missing on some perf though, my Gainward would boost to 2100Mhz but if you're not playing in 4k should be enough I guess

Your score @850mV is basically at 2070S level ...

Something was weird with my card, I got pwr limit. Tested with 1905@881mv, 16GHz dram and 105% pwr limit and score is now 7316, a bit too loud though.

 

[m6tzg6r]

On all the video cards i have owned i always put the Power Limit slider to max, but with my previous GTX 1070 i experimented lowering it, i noticed the performance would only slightly drop if i set it to 90% or even 80%, under 80% the performance loss became greater.

Now with all this recent talk about 30 series undervolting and the fact i recently got an RTX 3070 i tried lowering my Power Limit and going from 100% to 80% i reduced the average power consumption from 220W to 175W, which also reduced clock speed a bit, and also temp and fan noise. In game i lose about 3fps but i think it's worth it having a quieter system when gaming.

 

[Chrispy_]

On all the video cards i have owned i always put the Power Limit slider to max, but with my previous GTX 1070 i experimented lowering it, i noticed the performance would only slightly drop if i set it to 90% or even 80%, under 80% the performance loss became greater.

Now with all this recent talk about 30 series undervolting and the fact i recently got an RTX 3070 i tried lowering my Power Limit and going from 100% to 80% i reduced the average power consumption from 220W to 175W, which also reduced clock speed a bit, and also temp and fan noise. In game i lose about 3fps but i think it's worth it having a quieter system when gaming.

If you can be bothered to try different undervolted clocks, you can actually plot a curve of performance/Watt and there will be a sweet spot where you get the best framerate-to-noise ratio. I haven't bothered doing it for my 2070S but for the 5700XT it was way down at about 1450MHz, a good 25% less performance but also under half the power consumption. Given that I couldn't hear the fan at all even when it was running at 1666MHz, that was a profile I usually ran it on instead.

 

[Arcsaber]

If you can be bothered to try different undervolted clocks, you can actually plot a curve of performance/Watt and there will be a sweet spot where you get the best framerate-to-noise ratio. I haven't bothered doing it for my 2070S but for the 5700XT it was way down at about 1450MHz, a good 25% less performance but also under half the power consumption. Given that I couldn't hear the fan at all even when it was running at 1666MHz, that was a profile I usually ran it on instead.

I pretty much started at 100% and tested every 5% on the way down with OC Scanner until I got back to stock performance. That was between 75% and 80%, in my case.

 

[Taraquin]

I have updated the test with 105% powerlimit and rerun a few results due to get a more consistent score

 

[toilet pepper]

I recently got a 3080 Ventus and boy was it hot. Room temp was around 30c though but the card reaches 80c while gaming and the fans are not ramping up above 60% speed. I undervolted it to 1830mhz @ .818mv and a custom fan curve. It now barely reaches 70c with no discernable performance impact.

 

[nasky]

I have updated the test with 105% powerlimit and rerun a few results due to get a more consistent score

Seems the cooling is not that great on that card, for that undervolting temps should be lower thus reducing noise, I would replace thermal paste if there is no warranty stickers

 

[Taraquin]

Seems the cooling is not that great on that card, for that undervolting temps should be lower thus reducing noise, I would replace thermal paste if there is no warranty stickers

It's not great, but better than Ventus 2x and Zotac twin edge, but slightly worse than FE. The card is extremely quiet up to 40% fan rpm, 50% is okay, 60%+ is too noisy. I have not modified fan curve. Seems the card targets 65-70C. When pkaying Cyberpunk temp howers around 50-60C and 30% fan.

If you can be bothered to try different undervolted clocks, you can actually plot a curve of performance/Watt and there will be a sweet spot where you get the best framerate-to-noise ratio. I haven't bothered doing it for my 2070S but for the 5700XT it was way down at about 1450MHz, a good 25% less performance but also under half the power consumption. Given that I couldn't hear the fan at all even when it was running at 1666MHz, that was a profile I usually ran it on instead.

Powerlimit is very inefficient compared with curve editor. I got 50% lower powerconsumption with my 5700XT@1600MHz@850mv. Performance is 90% of stock so much better than pwr-limit. The yields of UV my 5700XT is better than the 3060ti as I get 92% perf@34% lower pwr consumption. Stock voltage at load with 5700XT is 1200mv (1150mv avg), 3060ti does a peak of 1080mv (950mv avg) so I basically lowered voltage by 300mv on 5700XT and 200mv on 3060ti.

 

[Chrispy_]

It's not great, but better than Ventus 2x and Zotac twin edge, but slightly worse than FE. The card is extremely quiet up to 40% fan rpm, 50% is okay, 60%+ is too noisy. I have not modified fan curve. Seems the card targets 65-70C. When pkaying Cyberpunk temp howers around 50-60C and 30% fan.


Powerlimit is very inefficient compared with curve editor. I got 50% lower powerconsumption with my 5700XT@1600MHz@850mv. Performance is 90% of stock so much better than pwr-limit. The yields of UV my 5700XT is better than the 3060ti as I get 92% perf@34% lower pwr consumption. Stock voltage at load with 5700XT is 1200mv (1150mv avg), 3060ti does a peak of 1080mv (950mv avg) so I basically lowered voltage by 300mv on 5700XT and 200mv on 3060ti.

Yeah, a lot of people don't realise that undervolting properly requires you to undervolt at different clocks so that you have enough results to define a new voltage curve.

Reduces the power limit does increase efficiency, but not as effectively as undervolting with a new voltage curve.

 

[Taraquin]

I updated with SOTTR-scores

 

[Jackrkl]

Hi, wanted to share my experience with UV my Palit 3060ti dual OC via Afterburner using Superosition benchmark at 1080p extreme and SOTTR@1080p highest with DX12 and TAA. I used the curve editor and powerlimiter. I updated the scores with the max 105% powerlimit and tested a bit more. The stated clockspeed is what I set it to in afterburner, BUT since the card has a stock OC of 30MHz I get real life frequency of 30MHz all the time in all games and Superposition.

Stock
6877 - 206W max - 80C max - fan@50%
-80% powerlimit (voltage stuck around 850mv, frequency around 1665MHz))
6504 - 165W max - 69C max - fan@50%
-1575MHz@700mv (useless as memoryclock gets stuck at 10GHz)
5562 - 119W max - 58C max - fan@30%
-1620@731mv (lowest voltage which allows for 14GHz dram clock)
6041 - 128W max - 65C max - fan@40%
-1665@750mv 105% powerlimit
6349 - 137W max - 67C max - fan@44%
-1800@806mv 105% powerlimit
6748 - 158W max - 65C max - fan@41%
-1905@881mv 105% powerlimit 16GHz dram
7316 - 208W max - 76C max - fan@70%


For low temp\noise\consumption 1665@750mv gives over 90% of performance for 40% lower consumption in Superposition and is inaudible and seems like the best setting for this as dram clock is kept at max, going down to 725mv dropped memclock by 4000MHz which plummets performance. 1800@850mv gives stock performance at 10% lower consumption, 3C lower temp and a bit less noise.

In SOTTR 1650@743mv gives you 88% of stock performance@ 7dB lower noise and 41% lower consumption. The OC-profile of 1920@881mv gives 3% more performance at 10% lower power consumption.

Thank you very much for this. I have a 3060ti FE and have been using this as a reference whilst I undervolt mine.

This is my first time undervolting. I understand the basics and altering things in MSI afterburner. But I wanted to ask is there anyway of recording the results like you have e.g. max wattage, fan speed etc
Or do you just sit watching the benchmark and write down the most common number ?
For some of the benchmarks I know they give you stats about avg FPS, max temp, 99th percent highest FPS etc

I am currently benchmarking in superposition and metro exodus. I am hoping to lower temps on my PC and give roughly the same performance as stock

 

[Taraquin]

Thank you very much for this. I have a 3060ti FE and have been using this as a reference whilst I undervolt mine.

This is my first time undervolting. I understand the basics and altering things in MSI afterburner. But I wanted to ask is there anyway of recording the results like you have e.g. max wattage, fan speed etc
Or do you just sit watching the benchmark and write down the most common number ?
For some of the benchmarks I know they give you stats about avg FPS, max temp, 99th percent highest FPS etc

I am currently benchmarking in superposition and metro exodus. I am hoping to lower temps on my PC and give roughly the same performance as stock

Unfirtunately I just observe powerdraw, but it is very consistent so my numbers are close to actual avg. As long as you are 100% gpu bound it will be quite accurate Max temp, powerdraw etc I use gpuz

 

[Taraquin]

A bit of follow up. It seems 30-series needs different voltages at different temps. In some games you can run 700mv but vram downclocks to 10GHz, in some games at a frequency line it will run the lowest voltage vram permits even at flat line. I set a flat line at 700mv@1530 vram at 15.6GHz on my 3060ti TUF and turned off fan, ran Hogwarts for 15mins and saw some thing interesting:
Below 30C: 1530@743mv
Below 40C: 1530@737mv
Below 48C: 1530@731mv
Below 56C: 1560@725mv
Below 65C: 1560@718mv
Below 70C: 1575@712mv here the game crashed due to too low voltage. I know it is stable at 10GHz vram though.

1590@731mv 15.6GHz with curve is always stable btw (frequency is 1620 once temp rises above 50C).

It seems vram vpltage requirements drops by 30mv going from 28C to 68C and core clock auto OC by 35MHz. Strange indeed, but interesting for geeks like me at 80C maybe 700mv would allow vram to run full speed? Not a temp I recommend though since hotspot then will be 90-95C.

On my old 2060 on a flat curve it runs at 712mv@1470 at below 40C, 706mv at 40-55C and 700mv at 55C+, but vram never downclocks and frequency stays the same.