The Math Behind GPU Power Consumption And PSUs

Power Consumption Measurement: Required Technology

With the graphics cards’ VR controllers switching back and forth so quickly, how much of this behavior is evident at the auxiliary power connectors? How difficult is it for the motherboard's slot and the PSU feeding those cables to keep pace?

First Steps With A Single Oscilloscope

HAMEG HMO 3054

View Deal

I planned the system around an emphasis on short measurement intervals. Unfortunately, this necessitated the use of a very good (and thus very expensive) oscilloscope with storage functionality. HAMEG’s HMO 3054 500MHz multi-channel oscilloscope served as the basis for my platform. It allows for the recording and storage of four channels, and has a helpful remote control via Ethernet.

The next ingredient I added came in the form of several high-resolution HAMEG HZO50 current probes, which can be used for these frequencies without a problem and, more importantly, can be operated without range switching.

Another important tool is a special riser card with loops, which has to feature extremely short signal paths to be fully compatible with modern third-generation PCIe graphics cards. This adapter is inserted between the motherboard and the cards themselves.

Those who think that the measurements could be taken through two rails on a 24-pin ATX power connector, looped through to the graphics card slots, would be wrong. All of the motherboard’s power comes from the same pool, after all. This simply can't be achieved without the riser.

With our arts and crafts project completed, it was time for the strategic partnership with HAMEG (Rohde & Schwarz) to shine. The project just can’t be completed without laboratory-grade technology and the experience needed to use it.

The preliminary result of my efforts, which you've seen in our reviews over the past few months, was a heartening success. However, even though four channels seem to be more than enough, a closer analysis reveals that we need more.

If the currents of both motherboard rails (3.3 and 12V) and the external PCIe power connector are measured, then all that’s left is one channel. It can be used to measure just one voltage. Since the PCIe connector sees the most action once the load goes up, I settled for taking that measurement and setting the voltages for the 3.3 and 12 V rails to those previously measured with a multimeter.

This approach reaches the limits of its usefulness once you get to graphics cards like AMD's Radeon R9 295X2. The PCIe connectors can’t be combined into one channel, since they serve different parts of the graphics card’s power supply. The currents can even flow the other way during short spikes, which I wasn't expecting to see.

Measuring with a single four-channel oscilloscope, three suitable current probe adapters and a probe might be in another league compared to measurements at the power outlet with a current clamp, or probes and multimeters. Unfortunately, it’s still not all that accurate. So what now?

Four-Point Measurements Times Two

Did I mention that this is an arms race of sorts? It sure feels that way if you’re trying to do this well! If the maximum four supply rails on a graphics card are supposed to be measured, then the oscilloscope needs to be able to handle a total of eight analog channels. After all, current and voltage need to be measured, and the results stored in parallel and in real time, at each of the four channels. Using previously measured values for the voltages can (and does) impact the results in a negative way, as we’ll see soon.

This is the point where I definitely had to make major changes to the measurement setup, and I once again enlisted our industry partner to help out. The solution consists of using two oscilloscopes in parallel (master/slave) to measure and record the data. The setup works perfectly at this point, allowing me to reliably measure a graphics card’s every last hiccup or sputter.

The bottom line is that we can measure currents and voltages down to a temporal resolution of a single microsecond. The interval needs to be adjusted depending on the application in question, because otherwise you might drown in the resulting massive amounts of data. For instance, if I conduct measurements for one-minute diagrams with a resolution of one millisecond, the oscilloscope averages the microsecond-resolution measurements for me.

We already talked about the measurement hardware in our previous foundation article. Here’s a quick overview of the most important information:

On the next page, we’ll get to what all of this looks like in detail and what conclusions can be drawn from it. The plot thickens, so read on!