Power Consumption

Cooling modern video cards is becoming more and more difficult, especially when users are asking for quiet cooling solutions. That's why the engineers are now paying much more attention to power consumption of new video card designs.

For this test we measure power consumption of only the graphics card, via PCI-Express power connector(s) and PCI-Express bus slot. A Keithley Integra 2700 with 6.5 digits is used for all measurements. Again, the values here reflect card only power consumption measured at DC VGA card inputs, not the whole system.

We chose 3DMark03 Nature as a standard test representing typical 3D usage because it offers: - very high power draw - high repeatability - is a standard benchmark that is supported by all cards - drivers are actively tested and optimized for it - supports all multi-GPU configurations - easy to obtain - fairly compact in size - test runs a constant duration and renders a non-static scene with variable complexity just like any normal game.

The four result values are as following:

• Idle: Windows Vista Aero sitting at the desktop (1280x1024 32-bit) all windows closed, drivers installed. Card left to warm up in idle until power draw is stable.
• Average: 3DMark03 Nature at 1280x1024, 6xAA, 16xAF. This results in the highest power consumption. Average of all readings (12 per second) while the test was rendering (no title screen).
• Peak: 3DMark03 Nature at 1280x1024, 6xAA, 16xAF. Highest single reading during the test.
• Maximum: Furmark Stability Test at 1280x1024, 0xAA. This results in a very high non-game power consumption that can typically be reached only with stress testing applications. Card left running stress test until power draw converged to a stable value.
• Blu-ray Playback: Power DVD 9 Ultra is used at a resolution of 1920x1200 to play back the Batman: The Dark Knight disc with GPU acceleration turned on. Playback starts around timecode 1:19 which has the highest data rates on the BD with up to 40 Mb/s. Playback left running until power draw converged to a stable value.

With the Radeon HD 6900 Series, AMD introduces a power draw limiting system. This serves the purpose to reduce power consumption from non-productive applications like Furmark or OCCT. As a result GPU designers will not have to worry about cards having issues with these extreme cases which promises to deliver cards that are better optimized for the general userbase.
NVIDIA has introduced such a system recently too, but both systems work on a fundamentally different level, each with advantages and drawbacks. NVIDIA's system measures actual card power draw via dedicated circuitry, AMD estimates it based GPU load percentage counters that watch the important functional units in the ASIC. Without the requirement for dedicated circuitry AMD's solution is more cost effective because it is a software solution. Software in this context means driver and mostly the SMC microcontroller that has been present inside AMD's GPU silicon for several generations doing clock, fan and thermal management, even though AMD claims changes were needed in Cayman for this system. AMD's system will only take into account the load numbers it gets from the GPU and guesstimates what the memory and power components will use, whereas NVIDIA's system can measure the whole card. Because AMD calculates the power consumption based on a formula this approach is independant of environment temperature, voltmodding and component tolerances - every card will show the exact same behaviour. NVIDIA has implemented application profiles which ensure that the limiting system will only be active in specific applications, AMD uses a general method that works without targeting specific applications but that means a higher (but still low) chance that real applications will be affected. AMD has added options to CCC for the user to adjust the power limit in two directions which helps the underclocking/power saving crowd and the enthusiasts; NVIDIA has no option for control of their power limiting system. AMD claims a sample rate "in the order of microseconds", I measured NVIDIA's GTX 570 to sample each 40 ms. I don't see a significant difference in either approach, maybe in theory AMD's system could provide slightly more stable clocks. But since such power limits should only be active in "useless" applications, it wouldn't make any difference, no ?

I did some testing of the feature with Furmark at several settings, as displayed in the second picture above. You can see that power draw and performance are very closely correlated for everything above the "-5%" setting, the relationship is quite linear. This shows that the system is working well and does what it is designed to do. Below -5% the results seems a bit more erratic, but still works somewhat well. Please note that the straight line at higher power levels means that Furmark can not load the card any higher and the card is running at maximum performance.

For our power consumption testing AMD's power control system was set to "0" for all tests. Only for Furmark Maximum, we set the slider to "+20".

Idle power consumption of the HD 6970 is on the upper end of the spectrum of modern graphics cards, but the difference is not significant enough to worry about it. Under load, power consumption is good, better than NVIDIA but not as good as the HD 6800 Series. My guess is that has to do with the new ultrafast GDDR5 memory they are using. We're seeing the same effect in Blu-ray power consumption where the result is horrible. 73 Watts for playing back a Blu-ray disc in 2010 is simply unacceptable. It is explained by the fact that AMD does not reduce their memory clock for media playback, so the GDDR5 memory is leeching tons of power at 1375 MHz.