Measuring Power Consumption
Measuring Power Draw
We were faced with the decision whether to measure the power draw of the entire test rig or just the individual cards. As measuring the entire rig isn't very transparent and doesn't yield comparable results, we chose to do it the hard way and measure the power dissipation of individual cards.
After conducting a lengthy series of measurements and test setups, we measured the power draw of our test rig without a graphics card at full CPU load. In order to minimize variance, we don’t use mechanical hard drives anymore, even though a disk-induced error would only be in the 2 or 3 W range. The PC is booted with a decades-old PCI graphics card, and the CPU maxed out with Prime95. Using this setup, we were able to determine a "baseline" draw of the system as 135 W.
In order to measure the power draw of our test cards as precisely as possible, we keep the PCI-based VGA card in the system, in addition to the card being tested, and subject the CPU to a Prime95 load at low priority while performing the actual test. This way, any remaining CPU capacity is consumed by Prime95 without influencing the actual benchmark. The baseline value is then subtracted from the system power draw of the current benchmark. The picture above shows that this works quite well. With a Radeon HD 7970 at idle, the system draws 150 W. Subtracting 135 W from that yields an idle power draw of a mere 15 W for the Radeon 7970, which matches the manufacturer’s spec.
One issue that needs to be discussed has to do with the power supply. The efficiency of a power supply isn't constant, but it changes with the load that the power supply delivers. If we measure the baseline power draw above at a certain PSU load, increasing the load during a benchmark will change the PSU's efficiency and cause the baseline load to be higher; the same setup as before now results in more power draw (due to the lower overall efficiency).
When we compared the power draw that we measured on the (secondary) DC side with the difference between the actual (primary) AC side wattage and the baseline wattage of 135 W, we got quite similar values. Determining the efficiency curve of the power supply and adding that to the equation, we can further refine this measurement and cross-check the accuracy of the clamp meter used for these tests. After tabulating all these measurements, the maximum deviation we ever observed was 5 W at a total system power draw of 435 W and, thus, a graphics card-only power draw of 300 W. As this is a deviation of less than two percent, we consider our method of measuring power draw accurate enough.