CPU Researching 'performance-per-watt' and power efficiency? (Undervolting, etc.)

[msystems]

I've done some experimenting with a gtx 1080 at 800mv (1.062v is stock), the boost had to be dropped down to 1720mhz which is about 200mhz less than the stock boost it was selecting with default voltage, however the TDP dropped to 62%. Another user hit around 1850 boost at 800mv (claims 68 watts usage) which means he had a pretty nice chip. https://www.reddit.com/r/nvidia/comments/666nnb/making_pascal_more_efficient_undervolting_and/

Even at this lower TDP I was surprised to find that 50% fan speed wasn't enough to keep it from eventually reaching the thermal throttle threshold. The fan gets quite annoying if it has to be ramped up above 40-50%

 

[zovc]

Well, I couldn't figure out anything about the 'overclocking' voltage and clock speed fluctuations at the upper voltages so I gave up on tuning those for now. My best guess is still that it's GPU Boost 3.0 and it ignores my locking of certain voltages and clock speeds? (This Galax Card's shroud says "GPU Boost 2.0"...) I spent some time working on that spreadsheet (Sheet 2, "0.825V") at the set voltage 0.825V and seeing how the card performed at different clock speeds.

Using Heaven as a benchmark versus actual gameplay means it's hard to appreciate the score we have for "performance," but it's definitely more efficient than recording 30+ minutes of gameplay... anyways.

I went from the highest stable clock speed I found, 1721Mhz (@0.825V) and got the score 4392, using an average of 100W. I kept working my way down and recording my results, I made it all the way down to 1531Mhz and got the score 4176, using an average of 96W.

There were no drastic or unusual fluctuations as I made my way down, though I did test several times at 1709 and my scores did vary a good bit. This tells me there's definitely a margin of error. The only way to get more reliable data is to run more tests at given voltages, and that's honestly a lot of time and effort that I'm not sure I can afford. We have these rough numbers, though!

So, I took the Heaven score and divided it by my average GPU power usage (in Watts), and this "performance per watt" number varied from ~42.5 to ~44.5, so nothing particularly drastic. Given my tests tonight (and recognizing there's a margin of error), 1645Mhz and 1683Mhz are the two most efficient clock speeds I tested at 0.825V.

I'm new to all of this stuff, so any suggestions or recording, organizing, or sharing this data would be helpful. Are there any certain ratios that would be useful? Are there any graphs I should be making?

 

[msystems]

There is one other thing, kind of beyond the scope of this which is that total system power consumption increases with temperature, so the ability to maintain the targeted power consumption (at the system level) depends on the efficiency of the cooling to keep the temperatures stable.
So simply pegging a voltage and clock isn't good enough- until the temperature completely stabilizes.


Pretty good thread here: https://forums.anandtech.com/thread...-power-consumption-with-the-i7-2600k.2200205/

 

[zovc]

That's a lot of info to digest (and I only glanced at it so far), but really interesting. It's definitely true that I haven't run the card long enough with a consistent synthetic to properly saturate its heatsink. So my temperatures are probably unstable... though they're also probably representative of a more 'consumer' load like gaming?

I'll try to put some thought into taking better control of my ambient temperature at the very least. My house's air conditioning and insulation aren't the best.

 

[Phryq]

Ok, after reading this, and likely not understanding.

What if these results are unique to each CPU? Each CPU overclocks differently, right?

And so maybe each CPU will be effected differently by under-volting?

 

[Biowarejak]

Right, so, even within the same series chip you'll have variations that enable them to perform better at different frequencies. AFAIK, the binning process is such that all chips from Intel start as an i7 and then, depending on what checks they fail, (or pass!) They get cores disabled and speeds adjusted.

 

[Thehack]

That's a lot of info to digest (and I only glanced at it so far), but really interesting. It's definitely true that I haven't run the card long enough with a consistent synthetic to properly saturate its heatsink. So my temperatures are probably unstable... though they're also probably representative of a more 'consumer' load like gaming?

I'll try to put some thought into taking better control of my ambient temperature at the very least. My house's air conditioning and insulation aren't the best.

It doesn't take long to saturate a heatsink. Probably 5 minutes is all it takes. If I run a gpu stress test it usually takes me only 3 minutes to hit stable temps. That said it is probably good practice to run at least 15 minutes to allow the whole case to heat up.

What you want to be careful is with water cooling. Water has a pretty high heat capacity (specific heat) - 4 times that of aluminum. So a 280mm radiator cooler will take a while to heat up. A lot of reviews will fail to test radiator cooling to saturate the heat capacity.

 

[zovc]

What if these results are unique to each CPU? Each CPU overclocks differently, right?

And so maybe each CPU will be effected differently by under-volting?

I could be mistaken, but I think it depends on what specifically you're doing. In general, the difference in how well parts overclock depends on the limits of what the chip is willing to be stressed into doing (in terms of heat and power usage) which just varies based on the physical differences of the parts. When it comes to undervolting, you aren't really stressing your parts so much as you are just giving them a little less. While--at a certain point--you are starting to "push" your chip as hard as it can go given the (lower) voltages you're feeding it, I believe it's different from doing that while feeding it as much voltage as it can possibly take.

I'm definitely not an expert on this stuff, so I'm just saying what makes intuitive sense to me rather than giving you hard science here. But, given that undervolting (and underclocking!) reduce thermals, I'm drawing the conclusion that a given architecture should perform pretty consistently, kind of like how any given i7-whatever should perform fine at its Stock settings.

It doesn't take long to saturate a heatsink. Probably 5 minutes is all it takes. If I run a gpu stress test it usually takes me only 3 minutes to hit stable temps. That said it is probably good practice to run at least 15 minutes to allow the whole case to heat up.

Kind of what I meant is, I didn't do much tuning at all. I just left my fan at 30% until the card started getting hot, then I locked it to 85% and was done with it.

It actually frustrates me a lot that I can't set my fans to speed up/slow down based on what my card is doing rather than how hot it currently is. Like, I wish there was a way to correlate my fan speeds to the amount of power the card was drawing or the clock speed it was running at instead of having my fans always be behind on cooling things down...

What I'm kind of trying to say is, it doesn't make sense to me for my card to say, "Oh darn! I'm 70C all of a sudden! I need to crank the fans up!" When I feel like we should have the technology for cards to say, "Hey, I'm using 70% of my power budget, so I'll definitely be heating up soon. I should increase my fans to [X]% while I'm drawing this much power!" Or a similar example based off of clock speeds.

 

[Phryq]

^^ Yes! That would really mean quieter fans and cooler CPUs in general.

I was thinking I'd just control fans manually. Turn the fan up before I'm gonna push things, and turn it down when I need quiet.