Discussion Feedback about GPU undervolting technique?

marcv81

Chassis Packer
Original poster
Jul 11, 2018
15
15
Hi team,

I am new to GPU undervolting and experimenting with MSI Afterburner on a GTX 1060. I'm mostly interested in reducing temperatures.

I understand that most people adjust the frequency/voltage curve. They keep the same max frequency, but lower the associated voltage. There are many videos on YouTube, I don't think I need to explain further; please let me know if that isn't the case.

I came up with a different technique, and would love to hear your feedback: I first settle on a stable overclock without changing the voltage, then I lower the power limit.

For instance my GTX 1060 has a 1949MHz default boost clock. I found out that it's stable up to 2050MHz. On the frequency/voltage curve I see the voltage at 2500MHz is 1081mV, and at 1949MHz it's 981mV: with this overclock I should be at the previous boost clock of 1949Mhz at 82.4% TDP (981^2/1081^2). So I set my power limit to 82% to prevent the card from overclocking and reach that frequency instead. In most benchmarks the scores and frequencies are almost exactly the same as before, but at lower temperatures and fan speeds.

So it looks like the method works perfectly, but there is more: the math above is a good approximation on average, but certain workloads draw more or less power than the average. For instance Furmark is known to draw a lot of power without reaching the max frequency. I believe the technique described has 2 advantages.
- It reduces the temperatures even for power-hungry loads like Furmark.
- It allows an overclock for loads which don't require much power. The performance should be slightly better in such cases.

I can confirm the first advantage in Furmark. The second one is theoretical at this point. EDIT: I can confirm the second advantage in GeekBench.

I hope my explanation is clear enough. Please let me know if you have any question and I'd love to hear what you all think!

Marc.
 
Last edited:
  • Like
Reactions: Arie

PLLovervoltage

Cable Smoosher
May 26, 2020
12
20
This is just simply playing with power limits. It should used in tandem with undervolting for best benefits. But just as you demonstrated, yes you'll see minimal performance decrease from reduction in power until a certain point. I've personally found from my own experimentation that the results I found reflected the same results the crypto mining community has. Which makes sense that they have a good grip on optimal power limits.
 

marcv81

Chassis Packer
Original poster
Jul 11, 2018
15
15
Thanks a lot for your reply @PLLovervoltage! Could you please explain the difference between the method I described and undervolting? Since I am reducing the voltage at each frequency on the curve I thought it did count as undervolting. But I'd love to hear the difference, and also about the benefits I am missing out on (to see if they outweigh those of my method).

Also thanks for mentioning cryptominers. I haven't read much about how they tune their hardware, I'll go check it out now!

Marc.