Undervolting AMD's Radeon R9 Fury For Better Efficiency
Voltages, Power Consumption And Temperatures
Why We Can't Actually "Undervolt"
Wouldn't it be great if you could simply adjust the GPU voltage? Well, PowerTune wouldn't be PowerTune if it made things that easy. Instead, it plays the arbitrator of the GPU's continued health by closely managing all things electric. This is probably one of the reasons that AMD had such a hard time with its decision to finally make the voltage target adjustable.
Taking a closer look at the voltage, the first thing we see is what appears to be a giant contradiction. Comparing the red curve, which represents the results at the default target, to the blue curve, which represents the results for the adjusted target, we see that the average voltage (VDDC) of 1.10V actually increases slightly to 1.12V! However, the 1.2V peaks drop to 1.1375V—extreme fluctuations in both directions are much smaller and the curve as a whole is smoother.
We also see two drops that show us our voltage reduction is taking the GPU very close to its limits. Going any lower would jeopardize the processor's stability. Even though the voltage increased slightly, the average strength of the electric current decreased significantly, resulting in lower power consumption.
So, how much does the smoother curve impact AMD GPU's power consumption? Let's take a look at the three graphs below. The first graph shows a 60-second window of the stock graphics card's power consumption. Right below is the load, nicely represented by the frame times.
We see that the card is running at maximum capacity all the time, and there are barely any differences between the scenes. Let's take one last look at frame times and then immediately shift our attention to the next graph.
The power consumption shows very clearly what's going on. Apart from the reduced power consumption, the peaks are much lower. If this was the whole story, we'd already be happy with the result. However, it also means that PowerTune is now able to react to changes in load and power demands much more precisely.
In the end, this pattern of results strongly suggests that we didn't actually undervolt, or permanently drop the voltage by a fixed offset at all. Instead, we increased the Power Estimation Engine's value in the telemetry's overall balancing act. The lower voltage number in MSI Afterburner just got rid of the reserve that AMD gave its early Fiji GPUs. At this point in time, the reserve seems to have outlived its usefulness.
Is The GPU Actually Cooler?
All of the above should mean that the GPU stays a lot cooler. There's a lot less waste heat, after all. Unfortunately, there's an enormous problem with this logic. The graphics card's arbitrator also controls the speed of the card's fans. Dependent on the firmware settings, AMD's graphics card recognizes the lower GPU temperatures and, consequently, doesn't spin the fans as fast.
That means the GPU actually ends up getting hotter. This isn't too much of a problem, since the Radeon R9 Fury Nitro's target temperature is set to 85 degrees Celsius. If you're wondering why the temperatures are going up, though, then take a look at how the arbitrator regulates the fan speed. Many different factors go into this. There’s the GPU temperature, of course, but there are also the currents reported by the VR chip. This is taken into consideration to make sure that the VRM and the rest of the board is cooled well, even if additional load is placed on the voltage regulators. So, the power draw is taken as an indicator for the amount of waste heat that’s generated.
Since our "undervolting" efforts are really just a decrease of the current, and with the GPU sensor weighted less heavily, our foray results in a less aggressive fan curve. The end result can be seen in this temperature graph:
If the card draws power like there's no tomorrow during our stress test, then its fans give their all to keep the GPU nice and frosty. As soon as it's "undervolted," the fans take it easy, the GPU runs warmer than it did before and the relevant parts of the board get just as hot. This counterintuitive outcome of the arbitrator's efforts also means that the card operates a lot more quietly, though.
At its hottest spot, the difference between the stock and the "undervolted" graphics card is less than one degree Celsius, which is basically within the measurement's margin of error.
Those who want a cooler graphics card, and not just one that's more quiet, need to manually adjust the fan curves.
Summary And Bottom Line
Lowering the voltage target really doesn't have any drawbacks. It's good for the electric bill, the environment and your ears. Even though it looks like we're undervolting, this setting really offers a way to adjust how PowerTune works. Voltage spikes can be reined in and overregulation dialed back. How far each individual GPU can go down this road before PowerTune can't keep it going any more needs to be ascertained on a case-by-case basis.
Our recommendation is to start at -48mV and run the card using this setting for a while. If everything works perfectly, then -72mV could be the next step down before trying to go for the maximum. Interestingly, this feature is now available for a number of AMD's other graphics cards as well, so their owners can also benefit from better efficiency. We'll test more boards and put our findings in an update or follow-up story.
MORE: Best Graphics Cards For The Money
MORE: All Graphics Content
MORE: Graphics Cards in the Forum
Igor Wallossek is a Senior Contributing Editor for Tom's Hardware Germany, covering CPUs and Graphics. Follow him on Twitter, Facebook and Google+.
Follow us on Facebook, Google+, RSS, Twitter and YouTube.
Please just show bar graphs or more normalized line graphs that compare multiple cards' power usage. I don't need to see the raw data guys.
I hope GPU vendors start getting smart (no pun intended) by making specfic models that automatically come with an undervolted GPU, but these GPUs would cost more since the vendors would have to take their best GPUs from production, test them, and then undervolt them. Similar to how EVGA picks out their GPUs with higher ASIC quality...but in this situation, it would be for power efficiency, not overclocking.
just wondering if any of the cards i have could benefit from this.
nVidia can afford to lose a few chips by binning. AMD cannot, they need to be able to sell all of the chips they can.
That said, nVidia probably does spend more time optimizing their uArch since all they handle is GPUs.
Interesting to see but it tells me that AMDs Radeon group needs to start kicking it up a notch to become more competitive. Power isn't the end all but it is pretty big when you have to consider the amount of extra heat your case now has to clear out.
The graph labeled "Gaming - Performance - UHD" may have incorrectly labeled the the graphics cards as "Sapphire Radeon R9 Fury Nitro" and "Sapphire's Radeon R9 Fury Nitro -48 mV." These are both AMD cards.
But the paragraph under that graph mentions that "Sapphire's Radeon R9 Fury Nitro does edge out Nvidia's offering," as if the graph meant to compare an AMD and an Nvidia card - not two AMD cards.
I think the label is wrong on the graph because the bars in the graph are red and green, suggesting an AMD-Nvidia comparison.
@Eggz:
This is a translation and the original text was a little bit different. Ok, It might be better to use other colors, you are right, but it was understandable here.
BTW:
Something was going wrong with this CMS system. I get another (empty) thread for this review:
http://www.tomshardware.co.uk/forum/id-2965949/undervolting-amd-radeon-fury-efficiency.html
I've tried also a MSI R9 390X Gaming 8GB. I got -100mV and 48 Watts less. The cheaper R9 380X can't be undervolted, because the VR chips are not adjustable. The voltage is mostly pin-coded and you need the specs to make a hard mod.
i undervolted the card and with the fan spinning a bit higher, i was able to get a 15c lower temp and get higher score in 3dmark because of no throttling.
Thanks for replying. Was it meant to be an AMD-AMD comparison or an AMD-Nvidia comparison?
".......The fact that AMD tends to set extremely high voltage targets for its graphics cards' GPUs can almost be considered standard at this point. Its strategy does maximize the number of GPUs that can be used, especially at the beginning of its production cycle, thereby increasing yield......"
Can someone elaborate please? Thank you in advance.
Someone else please correct me if I'm wrong, but I think the quoted language means to get at this: When a high voltage is applied to all cards, more cards will break, and people will need to buy new cards. If not, I think the major point is that (based on whatever motivation), AMD will sell more cards if it sets voltages high. The part about "especially at the beginning of its production cycle" could be a reference to the fact that AMD didn't originally allow voltages to be adjusted. From then until now, people who bought early faced a higher chance of their card dying from the voltage being too high. . . . This is based on interpretation and background knowledge, so it's just my best attempt to unpack the language you quoted. I'm open to other possible interpretations if there are any.
The minimum amount of voltage a chip needs to remain stable at a given speed varies from chip to chip due to tiny variations within the manufacturing process. A high target increases the amount of imperfections that a chip can have and still be stable (and thus can be sold) at the reference speed and voltage, at the cost of higher power draw and heat for the whole line.
"what other cards can take advantage of this? you state some others can but don't share which ones
just wondering if any of the cards i have could benefit from this."
I have a MSI R9 380 4GD5T OC and undervolted by 50 mV, and it's stable at stock clocks (980/1375 core/memory). Using default fan curve, I think fans stayed about 5-10% lower than at stock voltage when running Unigine Valley. Alternatively, I tried locking the fans at 50% (point at which they become quite noticeable for me), and temps dropped 5 C.
The charts label is wrong, it is already corrected in German version. Sorry for that, the review was written within two days
Nvidia is using the ASIC quality values to handle the voltage in their Boost feature more flexible. This is the reason why it makes no sense to undervolt a Nvidia GPU. AMD is using the same worst case voltage for all chips, to sell more working GPUs.
Could the Nano achieve 100w consumption?
I guess it's the same as with OC, more mature the tech (higher yields from the wafer) the more headroom, but still they don't bother to check it properly so instead they artificially limit performance (like ratio with CPU's, etc) despite the chip quality could be nearly the same across the whole line.
Anyways that's a good news meaning AMD is closing the gap to nVidia and possibly next iteration (Polaris) will live up to it's "chilly" name