We're still tracking the issue of Radeon R9 290X performance variance after our investigation into the matter last week and AMD's subsequent statement. As noted in that statement, AMD acknowledges that the apparent performance gap between the initial press samples and retail cards is wider than expected. Essentially, the variance from one 290X card to another ought not to be as broad as the 5-10% we're seeing. The firm is still investigating the reasons for this disparity, and an AMD rep paid a visit to Damage Labs yesterday in order to discuss the matter.
One thing the folks at AMD asked me to do is test the press and retail R9 290X cards against one another in the "uber" fan mode. Flipping the switch to enter "uber" mode raises the card's blower speed from 40% to 55% of its potential maximum. (Don't be fooled by the percentages there. The default 40% cap is loud, and the 55% cap is pretty brutal. At a full-on 100%, which you'd never experience during normal operation, the 290X blower sounds like a power tool. A noisy one.) You may recall from our original review that the 290X is quite a bit faster in "uber" mode. That's because in the default "quiet" mode, 290X cards constantly bump up against their thermal limits. "Uber" mode is intended to raise those limits by providing more cooling capacity.
At AMD's request, I ran our HIS retail 290X card and our original review sample through our 30-minute Skyrim test three times each and then through our MSI Kombustor worst-case torture test. The results are straightforward enough that I don't need to plot them for you. Cranking up the blower speed limit allows both the 290X press sample and the HIS retail card to run at a constant 1GHz in Skryim. There are virtually no clock speed reductions via PowerTune. Consequently, both cards perform the same in "uber" mode, averaging about 83 FPS during the duration of the test.
The results from Kombustor are similar. The press sample stays pretty much glued to 1GHz. The HIS retail card's GPU clock dips intermittently to as low as 978MHz in this peak thermal workload, but it spends the majority of its time at 1GHz, as well.
This is an important point to the folks at AMD, because it means the card-to-card performance variance we've cataloged can be eliminated quite literally at the flip of a switch. Owners of retail R9 290X cards will have to be willing to put up a non-default configuration that produces substantially more noise, but their cards should should be able to obtain the same performance that the 290X review sample achieved in "uber" mode in our review.
Obviously, the "uber" mode switch isn't a fix for everything that ails retail R9 290X cards. Many folks will prefer to have the combination of noise levels and performance offer by the stock configuration, and the card-to-card variance there remains an issue.
The next step from here is interesting, because AMD expects its partners to produce cards, like the DirectCU II 290X that Asus teased recently, with custom cooling that surpasses the stock cooler by a fair amount. With more effective cooling, these third-party cards could offer "uber" mode-type performance—and, potentially, less card-to-card variance—even at lower noise levels.
We'll have to see how that shakes out once we get our grubby little hands one of those custom 290X cards. I'm hoping that will happen soon. I also expect AMD to have something more detailed to say about the reasons for the unexpectedly broad card-to-card variance on current retail 290X cards before too long, so stay tuned.