Radeon R9 290X in windy weather for golden samples – SweClockers investigates

The approach is the same with almost every new launch. SweClockers and other media receive products directly from the manufacturers even before mass production begins, which admittedly makes it possible to be out early with reviews, but at the same time can raise some doubt as to whether the product in question really corresponds to what later appears on store shelves.

Usually this is not a problem, but sometimes suspicious cases arise where the media’s test specimens appear to be better than the products found at retailers. Some can be explained by natural variation, but unfortunately there are also so-called cherry picking or golden samples, where the best circuits are deliberately selected for reviews.

Rumors are now circulating that AMD may be dealing with the latter. According to a test by Tom’s Hardware, the media’s Radeon R9 290X performs significantly better than the graphics cards that the website has found at retailers, in some cases by as much as over 20 percent. This is far from what can be considered normal specimen variation or uncertainty in the measurements.

Time to investigate

To check how the country is, SweClockers is chasing the right to a Radeon R9 290X directly from dealers, which in this case is the Webhallen on Sveavägen in Stockholm. The copy in question is from Powercolor and the test method used is the same as in the original review of the Radeon R9 290X and later the Radeon R9 290, ie the card heats up for about ten minutes before data collection begins.

It may be worth mentioning that the Powercolor graphics card is completely identical to the standard reference model of the Radeon R9 290X, but comes in OC version with the graphics processor overclocked to the maximum frequency 1,030 MHz. To make the results comparable, SweClockers chooses to replace Powercolor’s firmware with the standard model, which limits the clock frequencies to 1,000 MHz GPU and 5,000 MHz GDDR5. The figures are also double-checked with traditional underclocking.

With the results in hand, the picture is clear. The performance is largely identical between the press copy and the graphics card from the store shelf, at least in the Über mode. A single frame per second differs, which is what may be considered normal variations or uncertainty in the measurements.

In the silent mode, where the dynamic frequencies have to work overtime, the situation becomes somewhat more murky. A certain difference in performance can be seen and even if it is not about large variations, the trend is noticeable. In the end, this means an average variance of a few percent, ie no extreme levels. The reason may, for example, be slightly poorer contact with the radiator, or simply slightly changing ambient temperature.

The conclusion is that there may be differences between the cards. SweClocker’s other copies, however, perform in line with expectations, and hopefully such extreme variations as Tom’s Hardware reports are unusual. The reviewer may very well have encountered a defective copy with, for example, a poorly fitted radiator, which on the other hand raises questions about the quality control.

However, it does not seem to be a general problem that affects all cards in the store, especially not for those who choose to use the performance mode Über, which is what AMD recommends for enthusiasts. SweClocker’s second Radeon R9 290X performs well within the limits of what can be considered reasonable.

However, the editors are not really putting the subject to rest yet, but are investing in running further tests tomorrow to further clarify the results. Stay tuned for updates!

Update 2013-11-06: Load for a long time

When the editors sensed a relatively small but still clear trend with the graphics card in Quiet Mode, the test rigs are loaded for a longer test. The goal is to find out if the difference between the press copy and the graphics card from Powercolor becomes larger over time when all components actually reach temperature.

To really test how the model behaves directly from the store shelf without any differences in software, original firmware from Powercolor with 3 percent overclocking (max. 1,030 MHz GPU) is used, not the standard variant as in the tests above. However, the behavior is also double-checked with the firmware of the press copy, and is identical except for the level at which the clock frequencies remain.

With the cards in Quiet Mode and under full load through Metro: Last Light for upwards of 60 minutes, the levels have time to stabilize. The difference in clock frequency is clear – the reference card from AMD is upwards of 100 MHz higher most of the time. The copy from Powercolor parks after a while at 748 MHz, and more or less refuses to move.

When the switch is switched to Über Mode, a completely different image is painted. Both units are able to maintain a maximum clock frequency, 1,000 MHz and 1,030 MHz, respectively, in principle uninterrupted even under load over time. The limit of 55 percent for the reference cooler fan simply seems to be enough.

Many small streams contribute to the differences

SweClockers has had long discussions with AMD during the day and has in the end waited to update the article while waiting for new information. AMD has read all the editors’ logs from the tests, and there are several theories about the behavior.

One of the most promising is about small variations in fan speed, which with the temperature-limited Quiet mode has a big impact. The fan of the press copy works at approximately 2,020 RPM at 40 percent, while Powercolor’s ditto “only” is around 1,920 RPM.

When SweClockers tests to manually set the maximum speed 42 percent, corresponding to about 2,050 RPM, the model from Powercolor performs noticeably better. The clock frequency is often above 800 MHz, and rarely or never goes down to what appears to be the base level at 748 MHz.

Another influencing factor is variations in the voltage of the graphics processor. Here, too, the differences are small but significant – Powercolor’s cards are often about 5 percent higher than the press copy. The result is extra heat, which in the limited Quiet mode makes a big difference.

The variations exist, but do not affect everyone

With another day of data on hand, it can be stated that there are differences between the graphics cards. The specimens from Powercolor have, among other things, a slightly slower fan in combination with a slightly higher voltage to the graphics processor, which together gives lower frequencies when the temperature stops.

However, this only affects the limited Quiet mode, where the fan speed is locked to a maximum of 40 percent. In more released Über, which is what SweClockers uses for reviews, the difference in performance is non-existent – both the press copy and Powercolor’s cards can work at peak frequency even for long periods.

AMD has a solution in the making

AMD is reportedly working on the problem, where the job-weary European department at the time of writing is inciting new awake colleagues in Canada. The company has a solution on the table, and promises one driver update within 24 hours which according to its own statement should minimize the variances.

We will be releasing a driver in the next 24 hours that corrects this behavior by normalizing all fan behavior. The 290X in Quiet mode should be at 2200RPM, and the 290 should be at 2650RPM.

In short, the company will normalize the fan speed of the various cards via software. At present, this is only controlled by a percentage, but after the update, the RPM value itself will also be taken into account. The company hopes that this will even out the differences between different copies, and at the same time takes the opportunity to lay the foundation for higher performance.

For the Radeon R9 290X, the value applies 2 200 RPM during Quiet Mode, which is comparable to approximately 1,920 RPM for the SweClockers Powercolor copy and 2,020 RPM for the press copy. This means a potential performance bump for both models – of course at the expense of a slightly higher noise level. The Über performance mode is left unchanged.

For scaled-down Radeon R9 290, the value is added 2 650 RPM, which can be compared to approximately 2,500 RPM for SweClocker’s press copies. This is also a higher value, which in theory means more performance. It is worth mentioning, however, that the fan speed for the Radeon R9 290 is already relatively good enough.

Update 2013-11-08: Catalyst 13.11 beta 9.2

As promised, AMD has launched new drivers, which by stabilizing the fan speed will even out any performance differences. The update applies to the Radeon R9 290X in Quiet Mode and the AMD Radeon R9 290, where the speeds are set at 2,200 RPM and 2,650 RPM respectively.

A quick look at the read values ​​shows that AMD’s adjustment works as intended. SweClocker’s press copy of the Radeon R9 290 X is around 2,200 RPM, just like the Powercolor card. The latter is also tested with the same firmware and frequencies as the press card, and even there it looks good.

The smoothed fan speed also means major changes for the clock frequencies. Where the difference between the Powercolor card and the press copy was previously around 10 percent or higher, it is now in the order of 4 percent.

The latter can be considered as normal variations for graphics cards with dynamic clock frequency, both in the case of AMD and Nvidia. As long as the levels depend on parameters such as temperature, seemingly identical cards will give slightly different values, as the quality of the circuits simply differs.

With this, SweClockers ends the monitoring around the performance deviations of the Radeon R9 290X. The editors will update the performance results for the Radeon R9 290X in Quiet Mode and the Radeon R9 290 for the next major article. Note again that Über Mode, where SweClockers focuses, is not affected by the changes.

Footnote: SweClockers would like to extend a big thank you to Webhallen, who against all odds managed to shake out a Powercolor Radeon R9 290X OC for this article.

Footnote 2: Feel free to read Emil Åkered’s post in the forum with further comments about the testing.