GPU undervolting around stupid limitations
I've spent the past few days playing with undervolting settings for my RTX 2060 Super.
TL;DR of the results: Around 15-20% performance loss vs stock, double that vs my usual overclocked profile. Power draw has gone down anywhere between 35-55%, depending on the benchmark. Synthetic benchmarks tend to have smaller power draw savings while games show bigger improvements. This is NOT a stable setting, more tweaking is needed. More details (and how to do this!) under the cut.
Just to be clear, this was an attempt to push voltage as low as possible with absolutely no regard for performance. I just wanted to see how low I could go while still technically running above stock clocks. I tried Steel Nomad and Time Spy for synthetic benchmarks, and a couple of games I frequently play for game tests. DLSS/FSR/XeSS were always OFF, all these tests are at native resolution.
-- IMPORTANT CAVEATS --
This is NOT a well controlled environment, this is me firing up games off my library, letting the temperature and framerate stabilise in a location that performs worse than average, then taking screenshots.
I do NOT have a good performance logging solution configured, nor do I have preset benchmark runs for these tests. They are quick and dirty and should be taken with more than one grain of salt.
The only thing that changed in between test runs was applying my custom profile on MSI Afterburner and returning to stock settings.
I expect the margin of error to be large, well above the 1-2% we usually see in professional benchmarks and reviews. I cannot verify exactly how much it is, so assume the worst.
Whenever "my overclocked profile", "overclocked" or anything with the same meaning is mentioned, it means the following settings in MSI Afterburner:
-- SYNTHETIC BENCHMARKS --
All tests run at default settings.
In 3DMark Steel Nomad, the card peaks at around 110 W for a decrease of 37% from stock, while performance drops by 13%. Compared to my overclocked profile, power draw went down by 41%, and performance went down by 18%. Results links: Undervolted (110 W), Stock (175 W), Overclocked (185 W). Note that the Undervolted run still has the VRAM overclocked for this test.
3DMark Time Spy had similar results. Performance loss was 14% vs stock and 17% vs overclocked, with power draw reduced by 43% and 46% respectively. Results links: Undervolted (100 W), Stock (175 W), Overclocked (185 W). Note that the Undervolted run still has the VRAM overclocked for this test.
-- GAME BENCHMARKS --
All settings and spots where screenshots were taken can be found here.
Assassin's Creed Mirage, custom Medium settings. The undervolted profile consistently draws 85 W or less even in more crowded areas. At stock settings the card draws between 175-180 W. That being said, this does come at the cost of significantly less stable frametimes, meaning this undervolt is probably on the absolute edge of stability and a bit more voltage would be necessary for a long-term setting. Note the over 20% performance loss alongside the nearly 500 MHz frequency drop.
Redout 2, max settings. Stock does incredibly well here, delivering over 100 FPS in most occasions. Tested at Neo-Tokyo, which tends to be one of the more performance-heavy tracks in my experience. Undervolting here dropped performance by up to 20%, and reduced power draw by 42%. Note that the overall frametime consistency has worsened here too, but it's fast enough that it doesn't detract from the experience.
Cyberpunk 2077, custom Medium settings. Stock settles in at around 60 FPS with my settings, while undervolting drags down to 50. Power draw drops below 100 W as a result.
-- HOW IT'S DONE --
Since Nvidia doesn't let you directly control the voltage of your GPU, you have to find a different way around it, and there is one:
Setting a maximum clock speed using MSI Afterburner's Curve Editor means your GPU will aim for that clock speed at the lowest voltage that matches it.
What this means is, if you make your V/F curve completely flat from one voltage onwards, your GPU will never try to go above that voltage. Thus, you effectively cap your card's voltage and select a target frequency for it.
For a visual example, here's my card's stock V/F curve:
And here's the undervolt setting:
There's a quick way to do this, so you don't have to set each little point manually. Select the leftmost point (lowest voltage) you want to try, set it to the frequency you want. Then, hold Shift, and drag across all the other points to select them together. Finally, hit Shift+Enter twice, and Afterburner will flatten the rest of the curve for you. Hit Apply and hope you don't crash.
-- THOUGHTS OVERALL --
Averaging out my results, power draw went down by 44% vs stock settings and 47% versus my overclocked profile. At the same time, the performance change was -18% versus stock. I didn't keep record of my overclocked gaming results, but the performance bump seemed to be around the same as that in synthetic benchmarks, if not a little less.
This makes for an efficiency improvement of 24% overall versus stock, accounting for performance loss versus power draw reduction. Not fucking bad, though again this DID really mess with frametime stability, so be mindful of that when trying this yourself. Frametime stability is just as (if not more) important than average framerates - A game running at 100 FPS with frequent stutters will usually not feel as good as a game running a rock-solid 75 or even 60, depending on the severity of said stutters.
My next attempt will be running stock clocks undervolted, though of course that's probably not going to be anywhere near as big an efficiency improvement. I wanted to go as hard as I could for the start, and 0.712 volts was what it settled down in.
