Are retail R9 290X boards slower than press samples, and does the AMD driver update fix it?
Last month AMD launched the Radeon R9 290X GPU, and overall it went very well for AMD as the card was heralded for its incredible price-to-performance ratio compared to Nvidia's top silicon. Shortly after the launch however, a few media outlets got ahold of some retail boards and found them to be much slower than the cards sent to them by AMD. Naturally, people suspected foul play, but AMD insisted it was just a driver issue, in that retail boards and press boards were using different fan speeds, thus delivering different levels of performance. It quickly issued a new driver with Catalyst 13.11 Beta 9.2 and we decided to test and see what the problem was, how the press board differed from the retail board, and whether AMD's latest drive resolved the issue.
So what is the issue?
When the Radeon R9 290X launched, we learned that AMD's Hawaii GPUs handle clock speeds a bit different than cards that have come before it. Instead of shipping with a guaranteed base clock speed, the Hawaii cards instead have a "maximum clock speed" that they can reach if they are running cool enough. You can also dictate what the card's maximum temperature is through the AMD PowerTune software, and once the card reaches that temperature it begins to throttle in order to stay below that threshold. On the Radeon R9 290X, by default it will run at 1,000MHz if it's nice and cool, but once it gets up to 94C it will begin to throttle since by default its max temp is 95C. So, at 94C it will begin to throttle, and will drop down to 940MHz, or 850MHz, or even lower if it's really stressed out. It's impossible to accurately predict with 100 percent certainty how an individual card will throttle, as all cards will behave slightly differently due to environmental factors as well as differences between individual chips, which is why people were curious about how other cards "in the wild" would behave.
In our tests, we found the R9 290X would throttle a tiny bit to around 900MHz or so once it reached 94C, which isn't too far off the mark. When the card throttles by 50-100MHz, performance isn't affected much. However, other media outlets were reporting that it throttled much more than that, even going as low as 727MHz and staying there for awhile. Nvidia also made this claim with a scary PowerPoint presentation prepared for the launch of the GTX 780 Ti. Tom's Hardware said the retail boards it purchased were slower than the less expensive R9 290 boards.
This caused AMD to release a statement to The Tech Report, stating, "A media outlet has uniquely reported instances of AMD Radeon R9 290X boards purchased in retail that have exhibited an uncharacteristic level of performance variance as compared to press samples issued by AMD. We're working to secure the board(s) in question for further analysis. Boards purchased by other media outlets have not exhibited similar characteristics that we're aware of. In the meantime, we've identified areas where variability can be minimized and are working on a driver update which will minimize this variance. We will provide an update shortly."
Shortly thereafter, AMD released a new driver named Catalyst 13.11 Beta 9.2 that promised to normalize fan speeds across all Radeons R9 290X GPUs.
Retail versus press boards
Our retail test board - a still shrink wrapped PowerColor R9 290X Overclocked.
Amidst all this hoopla we scored ourselves a retail board made by Powercolor (a third party sent us the board for testing). The card was still wrapped in plastic and had never been touched by human hands. To begin our tests, we ran all of our benchmarks with both the press sample Radeon R9 290X and the PowerColor using the original driver we used for testing -- Catalyst 13.11 Beta 9. Here are the benchmark results:
Radeon R9 290X benchmarks at 2560x1600 with the "original" Beta 9 driver
Best scores are bolded.
As you can see, just looking at this chart indicates there's a problem, as the press board is clearly faster than the retail card using the same driver. This was the same behavior indicated by reports from other media outlets, and Nvidia as well. However, by leaving the driver the same and swapping out the GPUs we were able to track down the source of the problem...
Click the next page to read our analysis on the problem.
The fan on the press board is set to run at a maximum of 40 percent under load, which in its default "quiet" mode is 2,125rpm.
The fan on the retail board is set to run at a maximum of 40 percent under load, which in its default "quiet mode is 2,025rpm.
We also independently verified the fan speeds using a digital tachometer. It required us to place a reflective sticker on the GPU's blower fan, and then a laser tracks how many times it sends back a signal during a short interval.
We also noticed on Legitreviews that Editor Nathan Kirsch recognized that his press board and a retail board were running different versions of the card's BIOS. We checked and sure enough our cards were also running different versions of what appears to be the same BIOS. The press board's BIOS was numbered "015.039.000.007" while the PowerColor board's BIOS was numbered "015.039.000.006," which appears to be one revision older.
When we asked AMD About the different BIOSes, it just said, "the variance was due to fan speeds, which we can control through the driver."
Here is that driver, which is the "Beta 9.2" driver designed to "minimize the variance."
Benchmarks at 2560x 1600 with the Catalyst 13.11 Beta 9.2 "fix" driver
Best scores are bolded.
As you can see, with the Beta 9.2 driver installed, the retail board achieves performance parity with our press board.
When we reached out to AMD and asked them to tell us specifically what caused the variance in fan speed between the retail boards and our press board, we received the following response. "The variance in production was larger than what we saw before launch." Apparently AMD set all the fans to run at "40 percent" and figured this would translate into the same fan speed for all GPUs. However, according to AMD, because of the way PowerTune fine-tunes voltages, fan speeds, and clock speeds, that 40 percent setting still leaves enough room for there to be differences in fan speeds, which translates to perfomance differences.
To corrrect this situation, AMD issued the new driver (13.11 Beta 9.2 and all versions beyond that) which normalized all GPU fans at 2,200rpm in Quiet Mode on the R9 290X. For the one retail board we have tested, the resolution totally worked.
Why does the variance only favor AMD?
The most curious aspect of this whole fiasco is that AMD says it's being caused by "variance," which it didn't see in pre-production. Now, variance would normally imply that you'd see some boards that are faster than the press board, and some that are slower, but in this case all of the retail boards tested by the media have been the same speed or slower than the press boards. In fact, we can't recall a single website that has found a retail board that is faster than one of the press review units. This does not appear to us as "variance" but more like AMD cherry-picking review samples. To be fair, we are certain Nvidia tests every GPU it sends to media as well, but we don't believe it tests a dozen cards and then sends the best one to a particular media outlet. We don't know if AMD did that for the R9 290X or not but it certainly appears that the review cards are as good as it gets.
To its credit, AMD is addressing these issues, or at least acknowledging them. In its latest statement to the Tech Report, it wrote, "These changes [in PowerTune] will also result in some degree of run-to-run test variability based on environmental and operational conditions in un-controlled test environments."
Regarding the question of why the variance only favors AMD, it wrote, "Reasonably we would expect the variability to occur both above and below the performance of the press samples, however it appears that most reported performances are biased towards the low side."
AMD concludes its statement by saying, "We are actively investigating these reports and we will update when we have completed our investigation."
At this point, it seems unlikely we will ever get to the bottom of this situation because the average website can't go out and buy a dozen GPUs to test. We also looked online to see if we could get a larger sample of boards and the R9 290X was sold out at every major online retailer. Whether the cards are sold out because AMD sent very few to the channel or if they are just super popular, we're not sure, but we don't recommend the R9 290X as the R9 290 is a much better deal since it delivers performance that is very close to its big brother for $150 less. In fact, nothing in this story changes or reflects our feelings on the R9 290X. It's a great card, but AMD made the R9 290's performance a bit too close to it in our opinion, making the 290X irrelevant.
That said, if AMD conducts an "investigation," we are very curious to see what it comes up with in terms of an answer. What is abundantly clear is that AMD's first attempt at marketing its GPUs using only the maximum clock speed isn't working out so well very far, so perhaps it'll rethink this approach in the future.
We're also getting very close to the release of R9 290X boards with custom PCBs and cooling solutions, so we are very curious to see how those bad boys measure up to the press boards as well.