Helluva time undervolting a 3080

Cafuddled · January 9, 2021

Hey Everyone, just wanted to post on my time undervolting and as a byproducts overclocking an Asus Tuf OC 3080. The big issues I've had with this card is proving out stability, as all my old methods are failing. Running OCCT with error check mode seems to no longer be accurate. I can play with the utilization in OCCT 3D Test to have the card sit at a power level or clockspeed I want to ensure stability for. This does seem to work at first with errors showing at a certain clock speed and then stepping down showing no errors on an hour test. I've also done the same at lower clock speeds with 100% utilization on the slider to make sure it's not just the way OCCT lowers the utilization.

I would then max out my power limit or clock speed using Heaven. This method can give you quite a bit of control. Say the settings you are using causes the power limit to force the down stepping of the clock speed. You can reduce settings or cap the framerate at any framerate you like to allow the power limit to max out at a given clock speed. Typically when you see an unstable clock speed you will see it pausing, stuttering or full on crashing the program.

Through a mixture of the two I was able to get a power curve that was looking extremely promising and stable through days of testing an many more games tested. I was at ease passing any benchmark I could find without any issues and running 3Dmark getting great stable and very high clock speeds (this is due to it's low power usage on these benchmarking apps). I even got the number 1 spot with my CPU and GPU combo, didn't crash or get errors in a day+ testing.

So I had what I thought was a great clock speed/voltage combination with great thermal's, I settled on 1980MHz at 0.906v (technically an undervolt for this card as 1980MHz is it's max but normally at 1.06v) which would max out the clock speed while still being under the 370watt limit on even the hardest hitting app you can find.

But then I loaded Cyberpunk 2077. Now before people say anything, this game runs 100% stable on my system when using default settings or later revised settings, with the power limit maxed out, so it's not the game. But what ended up happening once I loading the game was it crashed to desktop after a minute, with a power usage I seen on many other games, with the max clock speed I've seen on other games (not locking the max framerate yet on any games for testing reasons). I tested 5 more times with it doing the same thing within a handful of minutes, so I knew I had an issue. Retested with other games, Heaven and OCCT for hours and nothing. Back to the game and still going down with it not dragging more power than the tests I'm doing.

So next bumped all the voltages quite a bit which started lowering the clock speeds due to the power limits, which made it pointless and then ended up lowering the clock speed max. To get this stable I had to actually settle on 0.875v at 1905MHz as the power needed to get even a slight boost was simply not sensible, almost a 0.1v for a 45MHz bump... pointless. The reasons for this I feel is that Cyberpunk, unlike most other games and even OCCT with error checking on 3D and System Power usage tests is able to induce and unable to tolerate even the most infrequent errors to other games. There also maybe another argument to be made about stability when RTX is in use, maybe stable for none RTX and stable with RTX are two different values, but then Battlefield 5 was stable on prolonged gaming sessions, all be it an older RTX game, certain features of RTX maybe?

So with all that said, I feel that a large amount of clock speeds being reported online for the 3080 are simply not stable results and should not be used as even a good starting point for what to expect on your efforts. Make sure to test on Cyberpunk or possibly other RTX enabled games, do not rely on OCCT anymore as even things that OCCT would pass even at 100% utilization max for an hour test were causing Heaven to crash or freeze 5 minutes into the test. I've scoured the internet and I've not seen anyone speaking on this, just throwing around the same high clock speeds on tests that I was able to pass with ease before trying Cyberpunk.

Thanks for reading

Cafuddled · January 10, 2021

Turns out this is caused by DLSS, it utilizes the tensor cores which utilizes all of the chip that is typically not being utilized by any other game. It's actually very normal to have to drop your clock speeds or up your voltages when using DLSS.

Which brings to question, why does every overclokcing article and YouTube video for the 3000 series not even mention it, they're all going around giving overclocking advise like its 2015! This is a big deal, anyone who has a close the the bone overclock is going to have fails when they launch DLSS enabled games and start blaming the game and not this aspect of 3000 series overclocking. Can see it now "Every other game works fine, stupid games it's made by peasants the publishers found in the streets".

Mark Kaine · January 10, 2021

I mean, it was kind of obvious that's a cyberpunk problem, but you got a point, they should definitely mention this in overclocking guides etc... probably most just haven't figured that out yet.