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

zovc

King of Cable Management
Original poster
Jan 5, 2017
852
602
Hey!

I've been preoccupied with life stuff and ended up having to pay on taxes instead of getting a return like I was hoping for, so my projects are a little bit on-hold for the time being... I should be assembling something janky in either my S4 mini or my MC600 to tide me over until I can move things along more again.

The long-story-short inspiration for this post is I'm interested in living in a camper/camper conversion/'tiny home on wheels', but I'd like to be able to continue PC gaming...

Unfortunately, being conscientious about power consumption and still trying to achieve acceptable or better gaming performance don't exactly go hand-in-hand. There's not much information out there at all, actually. Usually "low end" parts get glossed over by reviewers (and stop showing up on charts) and "industrial" low-power parts are never run through gaming benchmarks with (the potential for) GPU bottlenecks removed.

I know cpubenchmark.com has its table for performance per watt, but it's not the most useful set of data. At least, not for someone looking to assemble a system... most of the parts that show up on there are mobile and/or soldered-on solutions and a lot of them are only on there because they literally use ~5W. It's not exactly a "proportionate" comparison when you're comparing a chip that uses 5W and a chip that uses 115W--but it is "fair." I also don't know or understand how multiple cores/threads factor into how chips are ranked on that site or even just that table. It's kind of an arbitrary number to me.

So, here are some questions I have:

  • Are there any sources out there that see how parts do when they're undervolted/underclocked?
  • Is it feasible to undervolt (and underclock) a desktop-grade CPU that's 65W or higher until it's using ~5W and is still stable? I understand speeds would definitely be gimped at this range, this is more of a theoretical question.
  • Are GPUs able to be undervolted/underclocked similarly to CPUs? Has anyone documented what the scaling is 'normally' like?
  • Would one expect a higher-end chip to perform better on lower voltage than a lower-end chip? What about a newer chip versus an older one? (However you can answer this would be helpful, even if it's contrasting numbers of cores/threads or comparing clock speeds or comparing architectures' differences.)
  • I haven't fooled with memory overclocking much at all--
    • Is there an appreciable increase in power draw when overclocking memory?
    • I'm under the impression that overclocking your CPU makes it harder(?) to achieve stable overclocks on memory, does underclocking have the opposite effect?
    • When overclocking memory, do heat spreaders or DIMM height (standard vs low-profile) matter for safe thermals? Would cooling be necessary in any particular case?
  • Is there an appreciable difference in power consumption between 2.5" SATA SSDs, SD card-style storage, M.2 SATA, or M.2 PCIe NVME drives?

As I mentioned earlier, my "fun fund" didn't get the windfall I was hoping it would from a tax return, so I'm not yet in a position to start testing chips and whatnot, but I'm very interested in generating and recording my own data out as best as I can.

Also, I don't yet have a realistic expectation for what is a feasible amount of power that can be stored in a camper's battery system. If you have any numbers that could help me roughly frame that it'd definitely be appreciated. But I can look into that on my own, I have and just don't remember.

Thanks!
 

Kmpkt

Innovation through Miniaturization
KMPKT
Feb 1, 2016
3,380
5,917
I'm actually planning to do a comparison of a 7700K vs 7700T in this regard where I step down the base clock of the 7700K in 100-200 MHz increments looking at a series of pretty straightforward benchmarks as well as power drawn from the wall and thermals. Ultimately I want to know if an under clocked K chip can match the T chip with respect to performance per watt or if the binning process really is that useful. Also I am very interested to know if there is an optimal under clock where you lose the least performance for the given wattage. I would also like to do a similar experiment using a GTX 1050ti, 1060 and 1080 sometime in the future. I figure that each of these components will have an ideal performance per watt point as well which is what I'd like to find.
 

zovc

King of Cable Management
Original poster
Jan 5, 2017
852
602
I only have a 4790K to play with at the moment (I also have an Atom board I got for a file server, but I don't know if I'll be able to play with its clocks and voltages), but I'm going to do what I can with it.

I also have a Galax 1070 mini which I notice you skipped over, but I'll probably play with it when I'm tinkering. I do have a 980 that I still haven't sold that I could fool around with, too.
 

Kmpkt

Innovation through Miniaturization
KMPKT
Feb 1, 2016
3,380
5,917
Yeah I don't have a 1070 and certainly couldn't justify buying one for the purposes of this testing. I'm mostly interested in seeing how low the wattage on the 1080 can go before you're at 1060 performance levels. As far as getting the power down to 5W or whatever, I'm sure there's going to be some sort of limiting factor before that happens.
 

Ceros_X

King of Cable Management
Mar 8, 2016
746
655
I get where you are coming from (I have been looking at tiny houses for about 6 months, out of curiosity) but I feEl most modern computers down clock when not being utilized and then ramp up when called upon (gaming, etc). While I am interested in the results, I think maybe your research would be better spent on solar / battery isolator, etc vs trying to get your desktop down to 5W. Is anything really as efficent as turning it off when not in use, etc?

@Kmpkt please post the results of your research, very interested in the processor work!
 
  • Like
Reactions: zovc

Thehack

Spatial Philosopher
Creator
Bronze Supporter
Mar 6, 2016
2,740
3,506
J-hackcompany.com
Some points to help you get started.

1. You can certainly downclock and undervolt a lot of chips but around 20W we start getting into ultra low voltage territory. A regular desktop chip cannot get into the 5w range but ultra low voltage certainly can.

2. You can certainly do the same for GPU. An AMD 470 can be downclock and undervolt to 75W and can perform better than a 1050 TI. A found that he could downclock an RX 460 to the performance level of a 750 TI and found it has power consumption of 40W which is very impressive. Many slow cores perform better per watt than fast few cores. The issue is a balance of silicon cost.

3. New chips are more power efficient than older chips. A higher end chip is usually binned better for voltage. But some of that is debatable. When you were able to overclock locked Intel chips users found that there was little difference between a locked and an unlocked chip. Likely Intel fab process is so refined that all their chips are good and it's actually just artificial market separation. None this is really surprising. But a lot of this data is hidden and obscured for profit reasons.

4. Memory overclocking has little effect on performance in games on the Intel side. AMD ryzen benefits from overclocking the RAM. RAM overclocking has been separated from CPU clock speed for a long time now.

5. There is certainly differences between models and types of SSD on the power side. Check anandtech for their reviews. The general conclusion though is the net effect is small unless you're a very mobile user or you work in Enterprise.
 

Thehack

Spatial Philosopher
Creator
Bronze Supporter
Mar 6, 2016
2,740
3,506
J-hackcompany.com
https://www.reddit.com/r/Amd/comments/56ncbi/sapphire_fury_nitro_undervolting_and_underclocking/

Really interesting one is here: Almost 970/290 levels of performance while being completely passive
https://www.reddit.com/r/Amd/comments/59sid2/experience_with_fiji_part_4_no_fans_on_this_fury/

https://www.reddit.com/r/Amd/comments/54dskp/experience_with_fiji_part_2_how_far_can_you/

Here's a good investigation by tomshardware. Notice the heavy penalty on power on the GTX 1060 going from 1500mhz to 2100 mhz almost doubled the power consumption, and really only gained about 20% performance.

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-1060-pascal,4679-7.html
 

alexep7

Cable-Tie Ninja
Jan 30, 2017
169
127
I'm actually planning to do a comparison of a 7700K vs 7700T in this regard where I step down the base clock of the 7700K in 100-200 MHz increments looking at a series of pretty straightforward benchmarks as well as power drawn from the wall and thermals. Ultimately I want to know if an under clocked K chip can match the T chip with respect to performance per watt or if the binning process really is that useful.
Please share your findings here when you do, I'm really interested in this as well!
 

zovc

King of Cable Management
Original poster
Jan 5, 2017
852
602
@Ceros_X, I hear you, but pulling ~200+ Watts for a few hours is a lot of power when you're running off batteries (as far as I can tell), if you look at "Gaming" laptops, most are lucky to see more than two hours of full-on GPU usage. Of course a camper would have a bigger system than a laptop, but it would have other things to theoretically power, too!

Thanks for the input and links, @Thehack, I'll try to digest that stuff later this evening.
 

Smanci

SFF Lingo Aficionado
Apr 21, 2017
126
160
Interesting topic and glad to see interested people.

I have a tradition of undervolting my cards to shave degrees off the temperatures without sacrificing performance or stability. Here's a quick example of a 970 at 100% power limit and afterwards at 80% / +130Mhz for core. I don't have a wall meter at hand so no idea how much less juice was consumed but seems like a significant drop. Fan speed was locked to minimum in both runs.


I also have VLP (18.75mm) DDR3 sticks clocked to 2200Mhz @ 1.5V. No issues there.
 
Last edited:

zovc

King of Cable Management
Original poster
Jan 5, 2017
852
602
That's really exciting to see, @Smanci! Not only did you turn your card's power limit down and lower your temperatures, but you were still able to squeeze a bit of extra performance out!

It's like a win-win-win.
 

Smanci

SFF Lingo Aficionado
Apr 21, 2017
126
160
The win-win-win makes it indefinitely more satisfyting than the I'm gonna fry the VRM-type of clocking. I really need that wall meter for some more info though :cool:
 
  • Like
Reactions: Biowarejak and zovc

Smanci

SFF Lingo Aficionado
Apr 21, 2017
126
160
Okey folks I've got my hands on this nice little device that measures power consumption. I've no idea how accurate it is but it was provided by the local library & city-owned power company and can also be found in EU stores like Clas Ohlson under the name PM-300.

This is for the whole system excluding monitor. I'm observing a ~30W drop without any loss of performance. The results are reproducible and the V450S power supply I have is supposed to be highly efficient even at low loads so there shouldn't be any efficiency-related inaccuracies.

 
Last edited:

zovc

King of Cable Management
Original poster
Jan 5, 2017
852
602
Heck yeah! That would make a significant difference over time on batteries! :cool:

I hooked this A1SRi-2758F system with a 2.5" HDD and a 8GB RAM stick up to my Kill-a-Watt meter and it tells me it pulls about 30W when being CPU benchmarked. It's using about 15-20W browsing the web right now. Dunno how much it matters, but it's also driving a 1080p monitor off the iGPU and has a mouse and keyboard hooked up via USB.
 
  • Like
Reactions: Biowarejak

zovc

King of Cable Management
Original poster
Jan 5, 2017
852
602
This video is split between reviewing the AMD 500 series cards (specifically the 580), but the last portion if it is focused around Chill and undervolting and underclocking with Watt Man.

(Oops!) Edit:
In World of Warcraft (which, we all know isn't especially demanding), his 580 was using about half the power it was before tuning...

Is there a similar tool to Watt Man or AMD's new Chill for Nvidia cards?
 
  • Like
Reactions: Biowarejak

Thehack

Spatial Philosopher
Creator
Bronze Supporter
Mar 6, 2016
2,740
3,506
J-hackcompany.com
This video is split between reviewing the AMD 500 series cards (specifically the 580), but the last portion if it is focused around Chill and undervolting and underclocking with Watt Man.

(Oops!) Edit:
In World of Warcraft (which, we all know isn't especially demanding), his 580 was using about half the power it was before tuning...

Is there a similar tool to Watt Man or AMD's new Chill for Nvidia cards?

Not really. Watt Man is AMD's first party utility and has some nice features like chill. For NVidia you got Afterburner and the Manufacturer's utility. Those utilities work decently well for undervolting. For NVidia, because of the way GPU boost works, you just want to reduce the power limit and the GPU takes care of the clockspeed itself.
 
  • Like
Reactions: Raxe and Biowarejak

Zeroth Alpha

Cable Smoosher
Jul 24, 2016
12
18
Not really. Watt Man is AMD's first party utility and has some nice features like chill. For NVidia you got Afterburner and the Manufacturer's utility. Those utilities work decently well for undervolting. For NVidia, because of the way GPU boost works, you just want to reduce the power limit and the GPU takes care of the clockspeed itself.
I would like to expand on this. GPU Boost 3.0 actually allows you to change the voltage-frequency curve which allows you to have higher frequencies at lower voltages, and actually lock it there if you'd like. Changing the power limit will only change where on the curve the GPU operates, not the curve itself. I personally run PrimeGrid (a number crunching software) on my GTX1060 at 1500mhz at 800mV. It's cool enough at full load that the fans never spin up from the minimum fan speed.
 
  • Like
Reactions: zovc and Biowarejak

Biowarejak

Maker of Awesome | User 1615
Silver Supporter
Mar 6, 2017
1,731
2,219
Here I thought that GPU boost 3.0 was only for optimizing games and getting drivers :)
 

zovc

King of Cable Management
Original poster
Jan 5, 2017
852
602
Where do I go to tune GPU Boost like you're talking about? Is it in the Nvidia control panel?
 
  • Like
Reactions: Biowarejak