Since June 2005 and the publication of our NVIDIA GeForce 7800 GTX test, in addition to performance of course, we have systematically measured the power consumption of new graphics cards coming onto the market.
In 2005, as now, this is a relatively simple reading to take as, instead of measuring the power consumption of the graphics card alone, we took a reading of the whole machine at the wall socket using a wattmeter. So as to factor in CPU consumption, that can vary according to graphics card performance, we use a CPU load in addtion to a GPU load.
While these results allow us to classify the graphics cards one against another, it doesn't give us the exact power consumption of the graphics card alone.
A new protocolObtaining this reading is fairly complex as graphics cards get power from two sources:
- The PCI-Express x16 port, which can supply up to 75W
- Additional PCI-Express connectors, which can supply 75W in 6-pin versions and 150W in 8-pin versions
Measuring consumption used via the additional PCI-Express power supply connectors is relatively simple as you can take a reading of the voltage directly on the graphics card PCB. Current can be measured with a clip-on ammeter.
With the PCI-Express port, things are more complicated but we've managed to put a protocol into place that allows us to measure the power drawn by the graphics card from this source, whether by the 3.3V (generally negligeable and just a few watts) or by the 12V.
We can now give you these readings. We have results for no less than 79 cards, of which 6 are also sold under another name. This makes for readings for a total of 73 cards in all! Almost 1500 intensity and voltage readings allow us to give you precise power consumption for four different scenarios:
- In 2D, displaying the Windows 7 desktop
- Playing an H.264 Blu-ray (Power DVD + Chapter 3 of I-Robot)
- In the 3DMark06 Pixel Shader test at 1920*1200
- In the Stability Test and Xtreme mode with FurMark at 1920*1200
We use two demanding 3D load tests, as depending on architectures, we obtained different power consumption readings for each. The results are therefore complementary.