PC Perspective Advanced Power Testing
For our power consumption tests on new GPUs, we previewed a new measurement method last year that we are finally taking advantage of. This isn’t an “at the wall” measurement with a Watts Up or something similar to that; instead we are measuring power directly from the graphics card itself.
How do we do it? Simple in theory but surprisingly difficult in practice, we are intercepting the power being sent through the PCI Express bus as well as the ATX power connectors before they go to the graphics card and are directly measuring power draw with a 10 kHz DAQ (data acquisition) device. A huge thanks goes to Allyn for getting the setup up and running. We built a PCI Express bridge that is tapped to measure both 12v and 3.3v power and built some Corsair power cables that measure the 12v coming through those as well.
Click for a larger version
Click for a larger version
The GeForce GTX 1070, GTX 970 and Radeon R9 Nano are well within reach of each when it comes to total power draw, but that yellow line representing the Radeon R9 390X and its 275 watt TDP is…higher. Clearly the Tahiti GPU is showing its age and its inability to compete with even AMD’s on Hawaii GPU in terms of power efficiency, and it definitely isn’t competing with the likes of the GTX 1070.
Another interesting data point from these results is in the comparison of the GTX 1070 and GTX 970. Even though the GTX 1070 is rated at 5 watts HIGHER than the GTX 970, our result show it actually uses less power (in our two tested games at least) than the Maxwell-based GM204 part.
Click for a larger version
Looking at only the NVIDIA products in our power testing scenarios, the GTX 980 Ti is the biggest draw here, as expected, with a 250 watt TDP.
Now that we are have direct power measurement numbers to use, rather than power from the wall results that could be tainted by increased CPU usage at any of these points, we can do some interesting performance per watt measurements for all parties involved.
Clearly the new 16nm GP104 is the most power efficient GPU we have ever tested, and the results above with the GTX 1070 compared to Maxwell, Hawaii and Tahiti prove that out.
Based on my observation of
Based on my observation of the GTX 970 and 980 releases, I have a feeling that the GTX 1070 will be the best value. And anyone who buys a GTX 1080 will regret it once the 1080 Ti’s releases. Personally I may end up getting just one 1080 just to try it out for gaming and folding@home, but I’m really eager to see what Nvidia brings to the table with the Titanium release.
The link on the “Testing
The link on the “Testing Suite and Methodology Update” page in this paragraph:
“For those of you that have never read about our Frame Rating capture-based performance analysis system, the following section is for you. If you have, feel free to jump straight into the benchmark action!!”
jumps to the 1080 review.
I was properly confused for a few seconds when I didn’t see any 1070 data on the page.
@Allyn: What would you think
@Allyn: What would you think about frame time weighted frame time percentile graphs? Like in the SSD reviews?
Just a joke, I don’t think it matters that much in this data since the variance is not multiple orders of magnitude here.
Ryan and I actually had this
Ryan and I actually had this conversation the other day. It could come into play with the percentile plots, but things would need to be presented a bit differently. It would help spread cards with greater variation out of the pack a bit more, but as it stands now, cards that misbehave tend to misbehave badly enough that we don't need to weigh it any differently to make it obvious.
These new power measurements
These new power measurements are amazing, thanks pcper for keeping on it, pushing measurement methods and supplying us with sensible data.
(however, I think the particular page mixes Hawaii, Fiji and Tahiti as others have also commented on)
“Testing suite” page:
>> As a
“Testing suite” page:
>> As a result, you’ll two sets of data in our benchmark pages
Word missing?
>> As a result, you’ll word
>> As a result, you’ll word missing two sets of data in our benchmark pages.
I know it’s already alot of
I know it’s already alot of work, but can we have some openCL or blender benchmarks? or even just from preimer pro testing
not all people game
and, well, also for the 1080 please! >.<
Using Chrome atm. When I
Using Chrome atm. When I click on a picture, the pictures tend to look a bit weird. Like, with the power graph when I click on it, the picture isn’t centered on the page. When I click on the bar graph, the picture is super large.
Would anyone be able to say
Would anyone be able to say if one could pair this GPU with a 980ti since they are comparable in performance and are pretty much the same architecture?
Unlikely nVidia would let you
Unlikely nVidia would let you do it. Might work in something like Ashes of the Singularity but betting other developers will do a similar version of multi-card rendering doesn’t seem like a sound plan.
Why single out power used by
Why single out power used by graphics card alone?
As long as GPUs need driver executed by CPU it does nor make sense to me.
Great review yet again Ryan.
Great review yet again Ryan. Just a heads up, the link to the benchmarks on page 3 sends one to the 1080 page.
Are the other cards used in
Are the other cards used in the comparison overclocked?
Why does this site still use
Why does this site still use the stupid tiny lines? Why can’t you just put the damn FPS numbers down and be done with it! I hate looking at very tiny lines just to get a idea of performance! This is a huge reason why I stopped coming to this site for reviews!
Ryan.. would you agree that
Ryan.. would you agree that nVidia probably made the 970 too good of a deal for what you got? As it seems there is more differences between the 1070 vs 1080 this time around.
If nVidia could change history, they probably would have either made the 970 not as fast or more expensive.
@Ryan Shrout, can u do
@Ryan Shrout, can u do another review regarding MICRON & SAMSUNG VRAM for GTX 1070 again?
there’s some fiasco like previous GTX 970 3.5GB VRAM & guess what now is bout the brand.
obviously, every reviewers cherry picked with SAMSUNG chip & how come there’s no MICRON chip for review??? thanks.