Radeon RX 480 Power Measurements And Possible Causes
[Update: On 7/5/2016, at 2:40 pm PT we updated a few sections on this page to reflect even more precise measurements that tell us that our original review analysis was correct regarding the load distribution across the card's six phases. This further clarity also resulted in removing some of our hypothesis text. We note below the sections we have updated.]
It’s not often that a graphics card launch is as polarizing as AMD’s Radeon RX 480 introduction. Our launch review produced graphs that have been interpreted incorrectly. Further, quotes were taken out of context, causing a great deal of debate. To set the record straight, we performed several follow-up tests, and are explaining the results in more depth.
In the meantime, AMD reacted to the issue through a statement, below, and we are going to comment on it later in the context of our findings.
As you know, we continuously tune our GPUs in order to maximize their performance within their given power envelopes and the speed of the memory interface, which in this case is an unprecedented 8Gbps for GDDR5. Recently, we identified select scenarios where the tuning of some RX 480 boards was not optimal. Fortunately, we can adjust the GPU's tuning via software in order to resolve this issue. We are already testing a driver that implements a fix, and we will provide an update to the community on our progress on Tuesday (July 5, 2016).
Some readers and colleagues have commented on our launch article’s measurements and graphs. They might even have a bit of a point when it comes to the latter, since it’s possible to draw the wrong conclusions without some background knowledge. Note that we’ve posted high-resolution graphs that are only one click away.
However, we use a low-pass filter and the highest sampling rate for our measurements to avoid aliasing effects and noise as much as possible. In addition, we’ve compared our measurement results to those of select graphics cards manufacturers over the years. Some of them use their own boards instead of oscilloscopes, but we’ve always ended up within the measurement error range with our results.
Note: Our power consumption measurement procedure and equipment is described in detail in our reference article on this very topic (The Math Behind GPU Power Consumption And PSUs).
In order to make things easier for a broader enthusiast audience, we increase the measurement intervals considerably and use new software for the analysis, which is able to evaluate very short load peaks (and valleys) and their timing. This procedure yields a much smoother curve, but it’s also a lot less irritating for readers. It goes without saying that we want to know when true spikes exist. However, that’s a story for another day.
Understanding The Terminology
Much virtual ink has been spilled on the topic of AMD’s Radeon RX 480 exceeding power targets and specifications. Before we can really get into this topic, it’s important to understand that this problem has nothing whatsoever to do with wattage. The only thing that matters is current. Consequently, we’re going over the Radeon RX 480's power supply and phase distribution first.
Without this overview, it’s not possible to make objective sense of the discussion on the specifications.
Replicating the Results
[Updated: 7/5/2016, 2:40 pm PT] A fact we got right in our Radeon RX 480 launch article: the card's six phases are distributed equally between the PCIe slot and the PCIe power connector. At first, we could only guess at the balance between all components, and our second time through we drew the wrong conclusions. This is why we decided to measure the connections of all phases ourselves. Based on this, we've also updated a small portion of the content of our hypothesis to reflect this.
We, and some of our colleagues, initially repeated our measurements because there were many questions about our original results. This time, we lowered the temporal resolution and, in turn, increased the filtering. Overall, our new measurement results come in approximately 1W below the previous ones, which falls within the usual measurement error range.
We performed individual measurements for the PCIe slot on the motherboard and the 12V PCIe power connector, as well as the PCIe slots’ 3.3V line. We also added all three to obtain the full board power.
In addition, we analyzed the GPU-Z log file, which was created in parallel with our other power consumption measurements. GPU-Z provides a raw measure of the GPU’s power consumption.
GPU-Z’s smallest possible measurement intervals are 500 ms, which is considerably longer than our measurement equipment’s. In order to add the GPU-Z results to our Excel chart, we used linear interpolation to create the missing data points. This might not be the most precise solution, but it should get the job done, especially since GPU-Z’s software-based, real-time analysis doesn’t exactly line up with the other measurements’ timeline anyway.
Beyond The 110W GPU
AMD is right when it says that the Radeon RX 480 GPU is a true 110W GPU. The average GPU-Z measurement result comes in exactly at that point. But that's not the whole story. After all, there’s still the rest of the graphics card, including other components that consume power. And then there are power losses as well.
Let’s take a look at the Radeon RX 480 below. We marked how we think the voltage converters are connected to the power supply lines based on our new round of measurements.
According to our measurements (see graph above), there’s a 50W difference between the overall board power at 164W and the separately-logged GPU power at 110W, if the PCIe-slot’s 3.3V connector is taken out of the equation from the start. The million dollar question is where exactly these 50W go.
Memory Power Consumption
The memory is an obvious place for speculation. Unfortunately, we weren’t able to get any exact power consumption numbers from Samsung. The only numbers that we do have are older and for the 4 Gb/s memory modules that consume 4.35W per gigabyte.
[Updated: 7/5/2016, 2:40 pm PT] There have been technological advances in the meantime and, in combination with some unsubstantiated statements, we’re thinking that we must calculate using 20W, or just slightly more.
Power Loss in the Power Converters and Other Components
It’s time to calculate the estimated power consumption for each of the six phases that serve the GPU. Distributing the GPU’s 110W evenly across its six phases leaves us with a power consumption of 18.3W per phase. In addition, there’s the power loss in the power converters, which amounts to 2.2W per phase.
This means that the GPU’s power consumption and the power loss in the power converters add up to approximately 20.5W per phase. These are all just estimates, of course, but we do feel that they’re plausible enough to draw some conclusions.
Distribution Across Power Supply Lines
[Updated: 7/5/2016, 2:40 pm PT] We’ve reached the most important part: the load distribution. Having figured out the distribution of the different loads, we can logically deduce that three of the six phases serving the GPU are connected to the PCIe slot, and three plus a separate phase on the top left are connected to the PCIe power connector. The phase for the memory modules is also connected to the PCIe slot (in the upper right, next to the PCIe connector).
Let's look at what this configuration means for the load distribution across all of the power supply lines:
|Motherboard PCIe Slot||Load||PCIe Power Connector||Load|
|3 of 6 Phases with 20.5W each (plus 1 extra at 20)||Approx. 82W (12V)||3 of 6 Phases with 20.5W each||Approx. 62W|
|1 Phase with 4W||4W (3.3V)||1 Phase with 16W||Approx 16W|
International Rectifier's IR 3567B could distribute the load between the phases asymmetrically instead, of course. But that wouldn’t make any sense, especially in light of the potentially unpredictable consequences and the looming overload of the six-pin PCIe power connector.
[Updated: 7/5/2016, 2:40 pm PT] Conversely, this also means that any other configuration would require an eight-pin PCIe power connector, which is the exact alternative that we said we preferred in our launch article, because it would be technically superior.
In comparing our measurement results with our hypothetical numbers, we find that our model fits both rounds of measurements well. On the next page, we finally take a look at the power consumption measurement results, discuss the supposedly exceeded specifications, and draw our conclusion.