I already posted a few observations regarding efficiency of NVIDIA GTX 1080 and 1070 at different core frequencies in the PrimeGrid LLR Races Thread 2016. Here are some more numbers.
The GTX 1080 used here is a reference board but water-cooled (for fun, not for extreme OC'ing).
The GTX 1070 is a factory-overclocked partner board (Gainward Phoenix GS), air-cooled.
Both cards were sitting in the same PC with i7-6950X @ 4.0 GHz, each card in a PCIe v3.0 16-lane slot, Windows 7, NVIDIA driver 373.06.
Folding@Home
These results come from a few sessions with different card settings, each session spanning several days, with a random mixture of WU types. As you know, PPD varies a lot between different WUs at F@H, making it cumbersome to benchmark for F@H.
Memory frequencies were left unmodified at effective 4.5 GHz (1080) and 3.8 GHz (1070) in all sessions.
Power was not measured externally but simply taken from the board sensors. Thus these numbers do not include power usage of the host PC (e.g. consumption of the supporting CPU process), and they are only as precise as the board sensors.
Power target of the cards was left at 100 % in each of these sessions. I achieved the different frequencies in case of the 1080 simply by not running Firefox in the 1.61 GHz session, and by having a Firefox window sitting minimized in the two 1.82 GHz session. In case of the 1070, 2.01 GHz were achieved simply by letting the card run with factory OC parameters, and 1.90 GHz by setting a negative frequency offset in MSI Afterburner.
Conclusion so far:
The 1080 is not only a bit faster than the 1070 (surprise), but it is also more energy-efficient while being faster. Plus, absolute power consumption is the same, hence same cooling solutions would produce same noise. But are these advantages of the 1080 over the 1070 worth its dramatically higher cost of purchase?
Remark:
Unfortunately, this particular factory-OCed 1070 turned out to be unstable at 2.0 GHz with F@H load. It has been stable at 1.9 GHz so far. However, my primitive way of downclocking it by a frequency offset without lowering the power target looks to be detrimental to efficiency. See below.
Edit, February 13:
This GTX 1070 from Gainward is in fact stable at the factory overclock, and at a trivial manual overclock to 2.1 GHz as well. I had underestimated the power draw of OC'd CPU + 2 GPUs; fixed this by a beefier PSU.
PrimeGrid
These are the same numbers as previously posted, only the 1080@1.8GHz row is new. All of these results are from "PPS (Sieve) 1.39 (cudaPPSsieve)" tasks.
Again, the power column shows numbers from board sensors, not from external measurements. The cards were neither temperature-limited nor CPU-limited in these tests. Only one GPU task per card was running at the same time. Core frequency, time per task, and power were averaged over several tasks. Variance between tasks of the same session was low, very much unlike the variance of F@H WUs. Another important difference between PrimeGrid and Folding@Home is that the former gives credit linearly to the amount of work done, while the latter gives extra credit for quicker task completion.
Conclusion so far:
Out of the box, the 1080 is both faster and more energy-efficient as the factory-OCed 1070. By lowering the power target substantially, you can make a 1070 as efficient, or even more efficient, than a stock 1080 in PrimeGrid (but not in F@H). However, it looks borderline impossible to clock a 1070 so high that it approaches the absolute performance of a stock 1080 in PrimeGrid. (But you can accomplish this in F@H with a good 1070 specimen - perhaps because then these cards begin to be CPU-limited with their single-threaded supporting CPU process.)
Still, the performance and efficiency differences that I am seeing here do not quite justify the difference of purchase price between these cards, do they?
The GTX 1080 used here is a reference board but water-cooled (for fun, not for extreme OC'ing).
The GTX 1070 is a factory-overclocked partner board (Gainward Phoenix GS), air-cooled.
Both cards were sitting in the same PC with i7-6950X @ 4.0 GHz, each card in a PCIe v3.0 16-lane slot, Windows 7, NVIDIA driver 373.06.
Folding@Home
These results come from a few sessions with different card settings, each session spanning several days, with a random mixture of WU types. As you know, PPD varies a lot between different WUs at F@H, making it cumbersome to benchmark for F@H.
Code:
card core points/day average energy
frequency (PPD) power per point
----------------------------------------------------
1080 1.61 GHz 680'000 109 W 14 J
1080 1.82 GHz 770'000 152 W 17 J
1080 1.82 GHz 760'000 152 W 17 J
----------------------------------------------------
1070 1.90 GHz 630'000 147 W 20 J
1070 2.01 GHz 710'000 150 W 18 JMemory frequencies were left unmodified at effective 4.5 GHz (1080) and 3.8 GHz (1070) in all sessions.
Power was not measured externally but simply taken from the board sensors. Thus these numbers do not include power usage of the host PC (e.g. consumption of the supporting CPU process), and they are only as precise as the board sensors.
Power target of the cards was left at 100 % in each of these sessions. I achieved the different frequencies in case of the 1080 simply by not running Firefox in the 1.61 GHz session, and by having a Firefox window sitting minimized in the two 1.82 GHz session. In case of the 1070, 2.01 GHz were achieved simply by letting the card run with factory OC parameters, and 1.90 GHz by setting a negative frequency offset in MSI Afterburner.
Conclusion so far:
The 1080 is not only a bit faster than the 1070 (surprise), but it is also more energy-efficient while being faster. Plus, absolute power consumption is the same, hence same cooling solutions would produce same noise. But are these advantages of the 1080 over the 1070 worth its dramatically higher cost of purchase?
Remark:
Edit, February 13:
This GTX 1070 from Gainward is in fact stable at the factory overclock, and at a trivial manual overclock to 2.1 GHz as well. I had underestimated the power draw of OC'd CPU + 2 GPUs; fixed this by a beefier PSU.
PrimeGrid
These are the same numbers as previously posted, only the 1080@1.8GHz row is new. All of these results are from "PPS (Sieve) 1.39 (cudaPPSsieve)" tasks.
Code:
card power core time power energy comment
target frequency per task per task
----------------------------------------------------------------------------------------------
1080 none 1.6 GHz 292 s 125 W 37 kJ baseline
1080 none 1.8 GHz 264 s 173 W 46 kJ
----------------------------------------------------------------------------------------------
1070 46 % 1.6 GHz 386 s 80 W 31 kJ same core frequency as baseline
1070 65 % 1.8 GHz 348 s 105 W 37 kJ same energy per task as baseline
1070 77 % 1.9 GHz 334 s 125 W 42 kJ same power as baseline
1070 none 2.0 GHz 319 s 165 W 53 kJ same power target as baselineAgain, the power column shows numbers from board sensors, not from external measurements. The cards were neither temperature-limited nor CPU-limited in these tests. Only one GPU task per card was running at the same time. Core frequency, time per task, and power were averaged over several tasks. Variance between tasks of the same session was low, very much unlike the variance of F@H WUs. Another important difference between PrimeGrid and Folding@Home is that the former gives credit linearly to the amount of work done, while the latter gives extra credit for quicker task completion.
Conclusion so far:
Out of the box, the 1080 is both faster and more energy-efficient as the factory-OCed 1070. By lowering the power target substantially, you can make a 1070 as efficient, or even more efficient, than a stock 1080 in PrimeGrid (but not in F@H). However, it looks borderline impossible to clock a 1070 so high that it approaches the absolute performance of a stock 1080 in PrimeGrid. (But you can accomplish this in F@H with a good 1070 specimen - perhaps because then these cards begin to be CPU-limited with their single-threaded supporting CPU process.)
Still, the performance and efficiency differences that I am seeing here do not quite justify the difference of purchase price between these cards, do they?
Last edited: