The Math Behind GPU Power Consumption And PSUs

Graphics Cards And Their Power Consumption Tricks

Changing Times: Load Spikes

Graphics cards aren’t the reliable “consumers” that they were a few years ago. This is due to how AMD and Nvidia are trying to increase efficiency, including the tricks they're using to get there. To simplify the explanation, I'm choosing AMD and its PowerTune technology as our example. Nvidia’s GPU Boost encounters the same issues, since its implementation and consequences are very similar to AMD’s.

So why are AMD and Nvidia going to all of this trouble? The idea is to adjust the GPU’s power consumption in real time depending on dynamic measurements of actual demand. In that way, the hardware doesn’t consume any more power than it really needs, similar to not driving your car at the engine's redline. Instead, you just shift to a higher gear and cruise along.

My analogy might not be perfect, but it does point you in the right direction. With PowerTune, AMD’s created a very complex scheme with many moving parts that keep influencing each other. The same goes for Nvidia’s GPU Boost technology. So, how does this really work if you go all the way down to the proverbial nuts and bolts?

Telemetry, Like In Formula 1

AMD’s PowerTune always starts by estimating the power consumption in real time, querying the thermal sensors and accounting for the telemetry data received from the voltage regulators. These values are transmitted to the pre-programmed power management arbitrator.

This arbitrator knows the power, temperature and current limits of the GPU (BIOS, driver). It controls all voltages, clock rates and fan speeds within these limits to try to maximize the graphics card’s performance. If even one of the settings is exceeded, the arbitrator can reduce the voltage or frequency.

The basic idea is to adjust operating parameters based on the actual power consumption needed in very short time intervals (nearly real-time) with the help of suitable voltage regulation circuits like the IR 3567B. AMD is currently on the second-generation SVI2 protocol, which is used by its most common ICs and APUs. These add and include control over the processor part via the northbridge.

But how short are the intervals in practice? In the past, there was a relatively long delay between the request for a higher voltage and the clock frequency adjustment based on it. AMD’s second-generation SVI2 moves the dot approximately two digits to the right, which is to say that it reacts in 10 µs instead of 1 ms. It also features much finer adjustments than its predecessor with 6.25mV steps.

Let’s take a look at a single short millisecond taken from an AMD Radeon R9 285 under load:

Unleash The Spikes: Load Spikes Like There’s No Tomorrow

If you’ve followed along so far, you might already have a bad feeling about what all of this means. A graphics card’s average power consumption really isn’t any different than a car's. If your vehicle has a real-time gauge integrated in its trip computer, and you looked down at it while driving around, then you know what I’m talking about. Some of the peaks can be massive, though they don't mean the trip computer is wrong. Their estimates are very accurate these days, and the massive peaks actually do exist.

The load peaks produced by the fast voltage changes pose a new set of challenges for high-performance graphics cards' power circuits. This also has some major implications for the design of PSUs.

The short introduction already shows that both AMD and Nvidia are able to stick to their self-prescribed TDP limits within very small tolerances. Should a measurement show that a graphics card is well above it, then either the measurement technology wasn’t up to the task or the graphics card is defective.

As we’ve seen in the past, sometimes a board partner will actually build in a “defect” (check out the circuit diagram below). We’ll go with what we’re most likely to encounter in practice.

Let’s put our heads together and come up with ways to measure this kind of fluctuating power consumption that allow for exact results and evaluation. It's a necessary step because the new technologies create a challenge for the reviewers interested in accurate measurements, and their readers who need to trust the data. Most traditional methodologies only provide a snapshot, which can be so far from reality that they're a biased estimate at best and, at their worst, completely wrong.

Any ideas for a good solution? Let’s begin our search on the next page.