Page 2:Gaming @ 1920x1080 (FHD)
Page 3:Gaming @ 3840x2160 (UHD)
Page 4:Lowering Consumption By Decreasing The Limit
Page 5:Efficiency For Different Applications
Page 6:Bottom Line
Page 7:Exploring Pump Noise
Page 8:Searching For The Actual Manufacturer
Page 9:Norms: Not the Norm?
Page 10:Cooler Master & AMD
Page 11:AMD’s Statement
Page 12:Information About The RMA
Page 13:Ending On A High Note
Page 14:Puzzled By Clueless Manufacturers
Page 15:Testing The Retail Card
Page 16:Spectrogram Frequency Analysis
Page 17:Do Dampening & Pressure Help?
Page 18:Results Overview
While AMD introduced its highly anticipated Radeon R9 Fury X in June, we’re taking a deeper look at its power consumption and pump noise.
Since we were able to procure our own retail board, we can finally take a much closer look at different aspects of the card’s power consumption. Our findings prove to be very interesting indeed! Furthermore, we’ve also taken several days to analyze the AMD Radeon R9 Fury X’s pump and the annoying noises that it produces. Even more interesting findings await!
We’re using the same benchmark system that we used for our AMD Radeon R9 Fury X launch article. However, we’ve tweaked it a bit. The idle power consumption’s now being measured using a system that has a lot of the software on it that tends to accumulate over time. Last time, the system was completely “fresh.” This time around, linear interpolation was applied to the data by the oscilloscope, and not when it was analyzed.
The biggest change, which we’ll stick with for all future launch articles as well, concerns the content, though.
We’ll look at power consumption in direct relation to gaming performance, and we’ll do so separately for 1920x1080 and 3840x2160, since there are major differences between the two. We’ll also look at several different games, and even run some of them with different settings, such as tessellation. In addition, we’ve added some applications that aren’t related to gaming. Life’s not always just about gaming, after all.
Bear in mind that these tests are merely snapshots in specific applications. A benchmark not part of our suite will likely fall somewhere between our average figure and the torture test peak. There are no absolutes these days; best estimates will always consist of a range.
Gaming @ 1920x1080 (FHD)
We know from AMD Radeon R9 Fury X 4GB Review that the Fury X doesn't do as well as many enthusiasts would have hoped compared to Nvidia's GeForce GTX 980 Ti. The averages of all of our results are very similar. At first glance, AMD’s 202W result doesn't look any worse than Nvidia's outcome (incidentally, also 202W). If gaming performance per watt is taken into account, though, then there’s a gap in Nvidia’s favor. But this doesn’t tell the whole story.
Flipping through the graphs, the individual results for power consumption (chart two) and the corresponding benchmarks (chart three) stick out. The insight gained from their comparison is quite simple: the reason that power consumption isn’t higher is that the graphics card limits it. This effect is particularly pronounced if the game involves tessellation. Metro: Last Light shows very clearly that tessellation is the Radeon R9 Fury X’s primary performance killer. What’s even more shocking is that Fiji's power consumption jumps way up with tessellation, whereas GM200 isn't as affected.
Gaming @ 3840x2160 (UHD)
UHD poses much more of a challenge for both graphics cards. Consequently, their power consumption increases significantly. On average, the scoreboard now shows 255W (283W peak in Metro: Last Light) for AMD’s offering and a much more reasonable 220W (233W peak in Metro: Last Light) for Nvidia’s card. The Radeon's power consumption does increase disproportionately, but its performance approaches that of Nvidia’s GeForce GTX 980 Ti as well. This means that efficiency takes a big hit. At least the performance is competitive, as can be seen in the benchmark results (chart three).
All of this means that the GeForce GTX 980 Ti is forcibly held back by GPU Boost and the restrictive 250W power target, but really doesn’t need any more than this to maintain the performance crown by a slight margin. Things are very different for the Radeon R9 Fury X, which occasionally draws up to 350W, even though few games hit the extreme reaches of more than 300W. Overall, there are two different philosophies at work here: GPU Boost’s hard limit versus PowerTune’s more generous allocation.
Lowering Consumption By Decreasing The Limit
What would happen if the power limit (not to be confused with Nvidia’s power target) was lowered manually? Would it be possible to trade a small and expected performance hit for significantly lower power consumption akin to what we see at Full HD? Unfortunately, no.
Check out the following graph. The y-axis shows the W/FPS ratio, which it to say that it shows how many watts are needed per frame per second. The x-axis shows the power limits. The numbers next to the lines stand for the actual FPS at the power target in question.
Even though the power consumption decreases from 267W to 170W when the power limit is set to -50 percent, the resulting frame rates just aren’t in the playable range any more. Higher power limit settings do not show imported results, either. The minimum and average frames per second, as well as the power consumption, stay the same. Things actually look about the same at Full HD, but the power consumption is low enough in that scenario that there’s no point to changing the settings anyway.
Efficiency For Different Applications
Let's shift away from gaming for a moment. Four tests show that professional software needs to be measured on a title by title basis because driver optimization, and with it application performance, can vary wildly. Generalizing any results just isn’t possible. We’re still offering a quick performance snapshot here.
We set the Radeon R9 Fury X’s performance as 100 percent, since the different benchmarks have dissimilar scoring systems.
Not taking performance into account, the overall power consumption duel can be summarized like this:
We again see that one or two benchmarks aren’t enough to provide a fair evaluation. Things have gotten so complicated that we had to rethink what we’ve been doing, and we updated our procedures accordingly.
It’s hard to draw a definitive conclusion about the Radeon R9 Fury X knowing that AMD says upcoming drivers will tease more performance out of it. At this point, all we can say is that AMD’s new graphics card catches up to the GeForce GTX 980 Ti’s power consumption at Full HD, but can’t keep up when it comes to performance. At UHD, its performance is competitive, but its power consumption is much higher compared to Nvidia’s graphics card.
We’re including links to overviews of the results for all tested graphics cards below. The graphs include the new numbers and open in a separate window.
Exploring Pump Noise
Let’s go back in time a few days to right before the Radeon R9 Fury X’s launch. Our German team (the guys measuring thermals, power and acoustics) were only able to test the new card for a few hours due to a scarcity of press samples. But this short time period was enough to give us an earache. Our colleagues reported the same problems, which confirmed the existence of a whistling noise. In addition, there was some kind of buzz. The latter is a low-frequency, oscillating grinding noise. These noises combined to form annoying acoustic fireworks.
Since we weren’t the only ones who contacted AMD about this, all reviewers received a preemptive email that was supposed to calm and reassure people:
Very small batch?! We wanted to know just how small this batch really was, and got our hands on our own retail sample from Newegg, just as a customer would. Our model was made by XFX, and we'll say more about it in a small footnote at the end.
Apart from the Cooler Master label being somewhat off-center, our press sample and the retail board, along with their pumps, are identical. By now, there’s talk on some forums about other labels and even engraved Cooler Master logos, but we haven’t seen a card like that in the lab yet.
The retail card’s noise profile is different from that of the press sample in the sense that the XFX card doesn’t just have a whistling noise, but also a lower saw-like noise. Really, it sounds like a buzz saw with an oscillating out-of-balance blade. It even includes the saw’s motor noise. The graphics card’s obviously not as loud, but the acoustic profile is quite similar. Consequently, we’ve labeled this noise “buzz.”
We tried to trace the origin of the sound with a special microphone designed to measure structure-borne noise. This led us to where the small heat pipes are connected above the voltage converters. Even though these are designed poorly from a flow perspective (sharp bends, unnecessary decreases in diameter, possibly a rough inner surface), they were neutral as far as the noise was concerned.
This means that micro-bubbles aren’t the culprit. This conclusion is confirmed by the fact that decreasing the pump’s voltage to slow it down halves the whistle's frequency, but doesn’t make it any less loud. All that remains as a possible source is the pump itself.
Searching For The Actual Manufacturer
To get to the bottom of this, the first thing we needed to find out was who actually manufactured the pumps. There might be a huge Cooler Master label attached to them, but let’s be clear: Cooler Master buys their components. They don’t make them.
According to our information, the Taiwanese company AVC (Asia Vital Components) made the pumps. This is the same company that manufactures the Seidon and Nepton products. It's not a small player in the OEM field, either. In fact, it sees itself as the world market leader in some areas. In spite of a number of personal contacts, all we were greeted with was a wall of silence when we contacted them.
Norms: Not the Norm?
Almost all large OEMs have been using the quality assurance norm ISO 9001:2008 for years. Internationally, it’s the most common one. It specifies minimum requirements for quality assurance management systems that companies have to implement to meet customer requirements and provide sufficient product and service quality. This most definitely includes a decrease in error rates.
An additional factor is process orientation. This means that all major operational processes are included and checked continuously. This can and should lead to ongoing optimization. We haven’t been able to find any negative experiences by other AVC customers. In fact, what we heard from them was quite the opposite: products like Enermax's Liqtech have very small and quiet pumps, and the Seidon and Nepton pumps are also acceptable. At this point, it looked like Cooler Master might just have chosen the wrong product for its purpose.
But is that really the problem? We heard through the grapevine that there have been discussions about glue problems with the batch in question. We couldn’t figure out if the issue was the amount or consistency of the glue, which means that this information’s just too vague to make a big deal out of it.
This is a plausible explanation though, since it would explain the permanent noise, which would be due to remnants of glue inside the tubing and its connectors that could cause turbulence. It would also explain why the severity of the problem seems to be different in various people’s accounts. This wouldn’t be due to the customer’s varying sensitivity to noise, but simply due to quality fluctuations.
Cooler Master & AMD
If Cooler Master is basically the man in the middle by offering a product that is comprised of a number of components made by other companies, then it’s Cooler Master's responsibility to verify that each batch meets the required quality standards. This is especially true for the first one. It takes some additional effort to dedicate personnel to the quality assurance of a newly-started production line, but this should still be covered by the certification.
This very common flaw was either missed or ignored during the time leading up to the launch. Why this might have happened is unclear.
It shouldn’t be overlooked that parts made by the original OEM (AVC) are often sent directly to the final customer’s (AMD) next OEM (Sapphire/PC Partner) and are immediately included in the final assembly on the manual insertion line without being checked at the point of arrival. The previous OEM is trusted not to have made any mistakes. Unfortunately, it seems that neither Cooler Master nor PC Partner conducted any spot checks.
A short statement by AMD reached us in the final minutes before we posted this article. We had shared our measurement results and insights with them leading up to its publication. The statement confirmed our information that a problem with the glue was the source of the noise problem.
Information About The RMA
In the end, there are a number of people suffering due to this massive demonstration of incompetence to implement decent quality assurance. There’s AMD, its partners that just buy these graphics cards from AMD after having them labeled and packaged in China, as well as the stores and the customers.
Looking at the first of AMD’s emails that we quoted at the beginning, it’s pretty clear by now that the “very small batch” was, at best, an understatement. We’ve contacted some of AMD’s partners directly to ask them if they are aware of the problem and willing to do spot checks. We also wanted to know how they are handling cards that customers found to be defective and if they have stopped delivering affected cards to stores.
We’d first like to note that all of AMD’s partners told us the exact same thing. We’re not reporting their names, since this information was mainly given by the R&D departments of the companies in question, and there haven’t been, and most probably won’t be, any official statements. This isn’t much of a problem, since the main message was that all of the spot checks yielded graphics cards with the same pump problem, even though its severity varied. None of AMD’s partners are planning to return the cards directly to AMD at this point for a variety of reasons and to avoid ending up on AMD’s bad side.
The good news is that AMD will apparently reimburse its partners for any losses suffered due to customers actually returning their graphics cards. Is this a ploy to sell at least part of the affected stock, because some customers aren’t that sensitive to noise and others don’t want to go to the trouble of an RMA? This would limit the financial damages, of course. However, it might still lead to undesirable results due to the damages to AMD’s and its partners' images. It’s questionable if the financial gain is worth it.
Ending On A High Note
We’ve been able to ascertain that there will be AMD Radeon R9 Fury X graphics cards with quiet pumps. Ultimately, the problem was found and fixed. The new revision won’t be identifiable by just looking at the package, though. It also stands to reason that everybody will first try to get rid of their old defective cards before pushing out the new ones. AMD could really have helped this situation by putting its foot down at the first sign of trouble. Then again, taking the high road does have to be financially feasible first. In spite of everything having been cleared up, a bad aftertaste remains.
Puzzled By Clueless Manufacturers
We’d like to add one quick note about the retail card before we get to our actual test with its video on the next page. The package states that the card has 4GB of memory. Right behind this, two black strips of paper were glued to it. The strips were fairly loose, so we took them off. Underneath, it says GDDR5. GDDR5 with a full 4096-bit interface? Whoever designed this really didn’t have any clue about the technology in the card.
At least they caught it before it was too late. In addition, Mantle is explicitly advertised in spite of this being the Radeon R9 Fury X. Games almost exclusively use this API as nothing more than a fallback. Windows 10, and with it DirectX 12, is right around the corner, and AMD's own website emphasizes its compatibility with the Microsoft API.
Let’s get to the good stuff!
Testing The Retail Card
We’re using a high-quality supercardioid attachable microphone. The following results should be interpreted accordingly.
The first thing we hear in the video are blubbering noises from the lower radiator. These didn’t go away after three days of continuous use. Since we’ve measured water temperatures in (and sometimes exceeding) the 60-degree Celsius range, and seeing that there’s no reservoir, there has to be enough air in the system to keep the expanding water from popping the tubes or the pump. Overall, the system is designed pretty close to the limit.
Spectrogram Frequency Analysis
The noise that sounds a bit like a saw consists of a permanent “carrier signal” between 1.93 and 1.97KHz and an oscillating buzz that’s located a bit lower on the spectrum. It’s really annoying. Let’s take another look at the frequency analysis that was recorded with a high-quality measurement microphone from a distance of 30cm.
The deeper parts aren’t as pronounced as the whistling noise, but can still be easily made out on the spectrogram. The higher frequencies far exceed 20KHz.
Do Dampening & Pressure Help?
We tried to glue dampening material to the inside, where it also applied pressure to the pump’s housing. This helps, but only a little. It doesn’t really impact the whistling noise, but cuts down the buzz a bit.
We also found that the part of the noise we call buzz had subsided a bit after 70 to 80 hours of continuous operation. This speaks for it being due to a manufacturing defect. The whistling noise did decrease a little as well, but only marginally.
We used the same test setup that we used for our launch article to take our new measurements. This yielded the following results:
|Gaming Loop Press Sample||33.1 dB(A)||33.7 dB(A)||34.2 dB(A)|
|Gaming Loop Retail Card||34.5 dB(A)||33.8 dB(A)||35.2 dB(A)|
|Gaming Loop Retail Card|
(After 80 Hours of Operation)
|33.6 dB(A)||33.6 dB(A)||34.6 dB(A)|
|Torture Test Press Sample||33.5 dB(A)||34.6 dB(A)||35.6 dB(A)|
|Torture Test Retail Card||34.9 dB(A)||34.9 dB(A)||36.2 dB(A)|
|Torture Test Retail Card|
(After 80 Hours of Operation)
|34.2 dB(A)||34.8 dB(A)||35.8 dB(A)|
What advice can we give to our readers interested in an AMD Radeon R9 Fury X? Wait or play the lottery? It’s important that AMD has apparently guaranteed that it will pay for RMA-related losses and has started doing so. The partners certainly hope that not all of the graphics cards end up back with them, though. People outside of Europe tend to care less about noise levels, which might explain why AMD chose to take this route. However, it’s not exactly great for the company's image.
The Radeon R9 Fury X and its cooling solution can’t be dismissed outright as an upgrade due to its acceptable performance. This wouldn’t be fair for the quiet cooling solution that might well be available soon.
From our point of view, Cooler Master and the OEMs it hired are responsible for this failure. Norms need to be adhered to and executed, especially if you proudly display certificates. We’re sure that AMD will get its money back from these companies, since these kinds of things are always subject to air-tight contract clauses in this business.