Extreme undervolting RTX 3070 Ti question.

pututu

pututu

[H]ard DC'er of the Year 2021
Joined
Dec 27, 2015
Messages
2,768
I've an EVGA RTX 3070 Ti XC3 GAMING with the latest firmware (94.04.5A.00.F9, driver 471.11) installed via Precision X1. I'm unable to reduce the gpu voltage below 818mV via MSI Afterburner (4.6.4 beta 3) nor with Precision X1. I'm reading somewhere that one can do 850mV with gpu clock around 1800-1900mhz. Below is the voltage-freq and gpu-z screenshot. I can't seem to force the gpu voltage lower than 818mV even though the gpu clock is already running quite low at 1455mhz. I can do this easily on 1080 ti running at 1450mhz at about 800mV a while ago. It can't be that pascal card is more power efficient than ampere. Maybe it is just specific to this model?

Anyone has any idea?

In case anyone ask why I'm doing this, I'm trying to figure out the ampere card power efficiency when running distributed computing, not gaming.



msi-png.png



3070ti-png.png
 
No reply? Someone must have an evga 3070 ti (or perhaps other model) and confirm if the card can be undervolted significantly.
 
Anyone with evga rtx 3070 ti xc3 or other models can confirm if extreme undervolting is possible on this 3070 ti?

Thx.
 
Yes, I did this before I setup the voltage-frequency curve. Just to be doubly sure, I also edited the MSIAfterburner.cfg file to allow voltage control and enter the overclocking EULA.
 
Have not tried undervolting, but just doing a -200 on my gpu clock and raising mem speed improved my mining performance from 46mh to 51mh and dropped power usage from 215w to 188w.
 
My other PC with evga rtx 3060 is able to undervolt to 787mV. I also upgraded the driver from 471.11 to 471.68 via DDU clean install on the 3070 ti but it is still stuck at 818mV. I can go higher voltage but not lower.

1630792087661.png
 
I think 818mV is the lowest mine will go but I'm not trying anything extreme.
 
Undervolting is purely silicon lottery. Some cards can do high clocks as low as 700mV, some require 800+ mV. I've seen 3080s undervolt to ~235W and still maintain stock performance, whereas others require at least 270W.
 
In this case it's probably a VBIOS limitation. You're probably out of luck.
 
I recommend overclocking your card to its highest stable clock then capping your power target around 83%. The performance below 83% seems to drop off rapidly on Ampere in my experience, but this might be different on a 3070Ti as I use a 3090, I would recommend atleast testing it though. It would likely be between 80-85% on the 3070Ti though.

This will put you in the cards optimal range atleast.
 
DFenz , thx for your feedback. I'm not into gaming (if that's what I think you are referring to) but into distributed computing something like mining but instead providing your free cpu/gpu computational power for research in maths, cosmology, physics, bio-medical science, etc. So likely that the computer will run for days, weeks and months continuously. I'm looking for the calculation (aka DC crunching) efficiency in terms of input power which is proportional to the square of the gpu voltage. I'm trying to duplicate some of what I did back in Dec 2016 but this time via controlling the voltage rather than the power limit.

Nvidia and AMD will continue to push for higher core clock and moar power. The next gen cards will be very power hungry base on unofficial leaks. Granted the consumer card is marketed for gaming but it has found its way into mining coin and into distributed computing (much less popular since there is little rewards when compare to mining or no rewards at all).
 
Shouldn't matter, technically it should apply to mining as well. But if you feel a locked voltage would be ideal thats the way to go.
 
pututu said:
DFenz , thx for your feedback. I'm not into gaming (if that's what I think you are referring to) but into distributed computing something like mining but instead providing your free cpu/gpu computational power for research in maths, cosmology, physics, bio-medical science, etc. So likely that the computer will run for days, weeks and months continuously. I'm looking for the calculation (aka DC crunching) efficiency in terms of input power which is proportional to the square of the gpu voltage. I'm trying to duplicate some of what I did back in Dec 2016 but this time via controlling the voltage rather than the power limit.

Nvidia and AMD will continue to push for higher core clock and moar power. The next gen cards will be very power hungry base on unofficial leaks. Granted the consumer card is marketed for gaming but it has found its way into mining coin and into distributed computing (much less popular since there is little rewards when compare to mining or no rewards at all).
Click to expand...

The PL method should work, but if you're dead set on voltage control, maybe try this: https://www.reddit.com/r/EtherMining/comments/8j2ur0/guide_how_to_use_nvidia_inspector_to_properly/
 
pututu said:
I've an EVGA RTX 3070 Ti XC3 GAMING with the latest firmware (94.04.5A.00.F9, driver 471.11) installed via Precision X1. I'm unable to reduce the gpu voltage below 818mV via MSI Afterburner (4.6.4 beta 3) nor with Precision X1. I'm reading somewhere that one can do 850mV with gpu clock around 1800-1900mhz. Below is the voltage-freq and gpu-z screenshot. I can't seem to force the gpu voltage lower than 818mV even though the gpu clock is already running quite low at 1455mhz. I can do this easily on 1080 ti running at 1450mhz at about 800mV a while ago. It can't be that pascal card is more power efficient than ampere. Maybe it is just specific to this model?

Anyone has any idea?

In case anyone ask why I'm doing this, I'm trying to figure out the ampere card power efficiency when running distributed computing, not gaming.



View attachment 383811



View attachment 383812
Click to expand...
OK pututu, I gave this a shot with my 3785 card. Real quick test with a randomly selected voltage/clock combo of 800mV/1530MHz. Here is my result, successfully set and tested below 818mV:

UPDATE: I can set it to 793mV successfully, but 787mV setting results in the card automatically raising it's voltage to 818mV under load. So I basically have 4 steps below the 818mV point where it will allow undervolting before it reverts to the "safe" setting. I imagine that some cards vary and these 4 steps are reduced to 0 steps, thus 818mV is your lowest.

GPUZ_UVex.jpg


AB_UVex.jpg

Firestrike Ext 3070Ti 0800-1530 on 10700K.jpg
 
Last edited:
Perhaps 3785 (Ultra Gaming) has better silicon than 3783 (XC Gaming) which is what I have. Anyway, thanks for reporting back.
 
You must log in or register to reply here.
Back
Top