How exactly does Turing GPU's Power Limit work?

They make noise, too.

Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee

Post Reply
Japanese Capacitor
Posts: 95
Joined: Wed Oct 21, 2020 5:16 am
Location: Warszawa, Poland

How exactly does Turing GPU's Power Limit work?

Post by Japanese Capacitor » Thu Mar 18, 2021 12:43 am

Lately I was experimenting with decreasing it and got amazing results. Lowering my card Power Limit by 30% shaved more than ten degrees with slowing down the fans to more than satysfying speeds. For the price of shaving also not much fps. I measured it with two rounds of Time Spy with default PL and two with -30%. Fps achieved differed by 6%. After that I tried the same with Yakuza Kiwami 2. It's the most demanding game I have right now on my drive and I noticed before, it keeps the fps really constant. The same 6% dicrease measured with few times walking one street from one end to another and observing fps - I'm not professional GPU tester, but if results were the same as Time Spy test, I assume, that executed it correctly.

Is all of this amazing deal of massive reduction of temps and noise for 6% less fps or I lose something else? Afaik, we want our GPUs boost as high as possible, because it increases fps, but also we want it to keep these high clocks, because it means constant fps - feeling of smoothness - and that's where power, temperature and voltage limits disturb. Right? If I understand it correct, I just wonder if with lowering PL I just hit this smoothness or it stays the same, but my card operate achieving lower maximum fps. Or maybe something else :) ?

CA_Steve
Moderator
Posts: 7539
Joined: Thu Oct 06, 2005 4:36 am
Location: St. Louis, MO

Re: How exactly does Turing GPU's Power Limit work?

Post by CA_Steve » Thu Mar 18, 2021 8:25 am

A couple of things to consider.

1) Power curve:
- Power consumption is proportional to voltage ^2 x frequency.
- CPUs and GPUs follow a programmed stepping: want freq x? Set the voltage to y. This happens on a dynamic basis as the workload changes.
- Card mfgrs may opt to tune the cards to always meet the high end of what's possible on the frequency curve, so they may put their 'thumb on the scale' and add a little bit of delta voltage on top of what's predicted to be needed.

Add all this up, and you get a voltage vs frequency plot like the one below (courtesy of Techpowerup's MSI GTX1660 Super Gaming X review).
Techpowerup MSI GTX1660 Super Gaming X clock and voltage profile.jpg
The effect of this plot is shown in the attached spreadsheet. While the max frequency only drops 4.5%, the power drops to 76%.
msi gtx1660 super gaming x power profile.png
2) The other thing to consider is it's not all about max fps. Consistent frame times and minimum fps also matters for smooth game play.
You do not have the required permissions to view the files attached to this post.

Japanese Capacitor
Posts: 95
Joined: Wed Oct 21, 2020 5:16 am
Location: Warszawa, Poland

Re: How exactly does Turing GPU's Power Limit work?

Post by Japanese Capacitor » Thu Mar 18, 2021 9:19 am

Thank you for such detailed explanation, but I don't fully understand it :)

By reducing the power limit, I reduced maxiumum voltage, so frequency the card is able to reach? So I reduced maximum fps my card can provide. Is it right?

I know, that feeling of smootness in games is connected more to frametime than framerate, but it's the thing I can't understand, how it works. Isn't frametime connected to boost clock and how constant it is? There's also minimum fps, which I know I want to have as high as possible, because it's connected to mentioned smoothness. And I assume, that minimum fps is consequence of boost clock's drops. Does it work like that or how?

Generally I'm curious, what I just did by reducing the power limit. Did I cut myself maxium fps or it affected minimum fps, frametimes or anything else?

CA_Steve
Moderator
Posts: 7539
Joined: Thu Oct 06, 2005 4:36 am
Location: St. Louis, MO

Re: How exactly does Turing GPU's Power Limit work?

Post by CA_Steve » Thu Mar 18, 2021 10:38 am

Japanese Capacitor wrote:
Thu Mar 18, 2021 9:19 am
Thank you for such detailed explanation, but I don't fully understand it :)

By reducing the power limit, I reduced maxiumum voltage, so frequency the card is able to reach? So I reduced maximum fps my card can provide. Is it right?
Somewhere during the gpu chip's test, or the gfx card's test, or both , a lookup table was created mapping the voltage needed to reach a certain frequency. When the GPU decides it needs to run at a new freq, the corresponding voltage is applied while the clock generator sets the new freq.
Japanese Capacitor wrote:
Thu Mar 18, 2021 9:19 am
I know, that feeling of smootness in games is connected more to frametime than framerate, but it's the thing I can't understand, how it works. Isn't frametime connected to boost clock and how constant it is? There's also minimum fps, which I know I want to have as high as possible, because it's connected to mentioned smoothness. And I assume, that minimum fps is consequence of boost clock's drops. Does it work like that or how?

Generally I'm curious, what I just did by reducing the power limit. Did I cut myself maxium fps or it affected minimum fps, frametimes or anything else?
For the last question: yes, maybe, maybe. Here's one simplistic example (because this could go deep very fast). Let's say you are playing a game that consistently runs at an average of 140 fps and you are using a 60Hz monitor. Using frame time analysis, you see that the minimum fps is running around 100fps. You power limit the card and now ave fps is 100fps and minimum is 65fps. Will you see anything different on your 60Hz monitor? Will it be "less smooth"? Probably not.

Here's the originator and gold standard for frame time analysis and why it matters. Here's hoping Scott is doing well, because the Tech Report really suffered with his departure.

Japanese Capacitor
Posts: 95
Joined: Wed Oct 21, 2020 5:16 am
Location: Warszawa, Poland

Re: How exactly does Turing GPU's Power Limit work?

Post by Japanese Capacitor » Fri Mar 19, 2021 12:38 pm

Thank you for an article. I will test my card in more games and see, if I just feel the differene caused by PL. After reducing it this card is just below noise of anything other and I like sound of my computer better... Sounds from the speakers most of the time cover it even with default PL, but quieter is always better. But as I always say, pursuit of silence should end, when we start to sacrifice performance. But this tempted me :)

Post Reply