I have a slightly dumb question, but there are enough hardware experts on here that I figured it was worth a try.
I’ve got a 3080 which has been great. I’ve recently been playing Dying Light 2 – only a year late to the party! – and it plays great. However, I’ve noticed my graphics card fans are running overtime while I’m playing, and the card appears to be getting hot (it’s very hot to the touch, but I haven’t run any diagnostics).
Any idea what settings should be turned down to reduce the load on the GPU? Vsync, DLSS, Reflex are all off at this point, and it’s still pretty loud and toasty. Are there any specific details settings that crank up the load on the card? Playing at 1440p.
I suggest undervolting your card on a curve via MSI afterburner. I have mine going to 900mV and that dropped power utilization from 385w to 330w, substantially dropping power utilization and noise while only hitting performance like 3%. Google for instructions, lots of them both written and on YouTube.
You can watch utilization, heat, fanspeeds, and power usage in realtime via the GeForce experience overlay. Or RTSS from MSI afterburner I guess, although I’ve never used it.
I would think enabling Vsync and limiting to 60hz would also give your card a break. Turning on DLSS would also make it work less to get to those 60hz. So turning on DLSS and Vsync should help.
Definitely don’t vsync to 60Hz, that compromises gameplay and homeboy didn’t buy a 3080 to run games like a pleb. DLSS should be used at “quality” basically everywhere it’s available and never ever below that (unless digital foundry recommends it, I suppose). Doesn’t matter if the game runs at 4k165 native with frametimes straight as a ruler, DLSS quality will look the same (or better) and use less power.
So yes use DLSS, but also undervolt on a curve.
Oh, and check digital foundry for Alex’s recommended optimized settings for that specific game.
It’s less optimal but you can just slide the power limit percentage in Afterburner down from 100 to 90 or 85 or whatever as a quick fix. You lose a lot less than 10 or 15 percent performance.
That works too but sacrifices a lot more performance. Why not learn the correct way to do it, and undervolt on a curve? It isn’t particularly difficult.
I power limited for long running non gaming workloads, and even 70 percent limit was 2 or 3 percent lower performance, which for me was not worth optimizing. Though now that I type that out it might be an worth checking if I could go even lower I suppose, with real tweaks.
That depends on the game. If you look at utilization with the GFE overlay you’ll find that a powerful card like the 3080 is usually not fully loaded when playing videogames, particularly older or nonintensive ones. Dying Light 2 I would expect a full load at 1440p with RT or 4k without, certainly 4k native.
Are you sure about this? That has certainly not been my experience with capping the power limit of the 3000-series cards. Undervolting is the only thing (for me anyway) that meaningfully reduces power draw without also meaningfully reducing performance. I would run the benchmarks again as you might find you’re only getting 80% out of the card with a 70% power limit.
I found a lot of (confusing and contradictory) advice on how best to set up Afterburner. I ended up following instructions, which basically consisted of “cap the frequency at your default boost clock for all higher higher frequencies,” so my “curve” ends up looking like this:
Is that what you’d recommend, or would you recommend something different?
That’s a valid question. I don’t see much (the card has only been in the case for about a year), but cleaning out the entire case (including the GPU) is on the to do list for the weekend (the Noctua case fans have visible dust on them).
Yes, that looks about right for a 950mV curve. Depends on the silicon lottery but you may be able to go lower without impacting performance, you basically need to test. Mine did 900mV no problem but 850mV I had crashes in benchmarks.
Your card only boosts to 1800Mhz? I have my curve at 1900.
If it’s still loud you should also look at the fan curve, set it to ramp up and down more slowly to avoid annoyance and use lower fan speeds tolerating higher temperatures. Basically just need to mess around with it.
I have the fan curve in Afterburner basically set to keep the fan low until it hits about 90 degrees –
I think the issue is that it’s hitting that 90 degrees and the fans are going wild. I’d be reluctant to set the fan lower at that temp, though, since I obviously don’t want to cook the card. Does that fan curve look crazy?
I’d basically cap it such that at its loudest it doesn’t drive you crazy. I don’t let mine go higher than 80% but again that will depend on the card, they all have different fans and cooling.
Remember the GPU will try to use as much power as possible, generating as much heat as possible, until it hits its thermal limit around 82C. It’s fine to stay there forever, it’ll just clock a bit lower.
I just tried it with Stable Diffusion to check. I couldn’t really detect a difference until I got below 70%, before that it was like 4.35 versus 4.31 it/s. After that it starts falling fast.
I would guess if I happened to be maximizing the GPU it would fall off faster. For example I’m not even increasing the batch size usually (which would use more memory at once) because I’m just playing around and prefer to get a single image super quick.
Oh yeah, ML is a different beast. I am not surprised that a 70% power limit achieves similar timings in something like stable diffusion. I have a MSI afterburner especially for Tensorflow in my 3090 that pumps memory clock and reduces core clock.
A game in 1440p or 4k will likely work very differently in regards to power limit vs. undervolting.
