Yesterday, we detailed how troubles with AMD’s R9 290′s fan profile could lead the GPU to return slower-than-expected performance under load. This was a problem that reportedly struck both reviewers and certain retail cards and left multiple sites pulling their recommendations on the R9 290. According to AMD, the problem was caused by improper fan settings on retail cards and imprecise tuning that relied on power modulation (PWM) rather than a specified RPM speed. As a result, the company said, some retail cards that were supposed to be spinning at 47% were actually spinning as much as 350 RPM slower — thus causing the overheat issue. Last night the company pushed a new driver that would supposedly fix this by relying on RPM tuning to set fan speeds rather than PWM manipulation.
Justin Jaynes, an acquaintance of mine at Seeking Alpha, offered to contribute some results to our testing as well. He’d just ordered and received a Sapphire 100362SR. Armed with his retail card and the reference GPU AMD shipped us with a default driver, we set out to test the lay of the land. There’s good news, from multiple quarters — but first, our test results.
These tests were run with our AMD R9 290 reference card across six runs of Metro: Last Light. As of yesterday, default behavior at the 47% fan profile speed showed a noticeable drop-off by the end of the run. Now, that drop-off has vanished. The reason you can’t see the previous 55% bar is because the new 47% driver tests obscure it. We tested further, and ran the card through a full 12 runs of Metro: Last Light. The results were the same — no drop-off in performance, even across a full 12 runs.
GPU-Z’s logged results were rather interesting. The fan speeds were a tad higher at each percentage — with the original driver, the R9 290′s 47% fan speed topped out around 2520 RPM. With the new driver, the 47% mark was a bit higher, at 2575 RPM. The difference in clock throttling, however, wasn’t caused by an RPM shift that small, but because according to GPU-Z, the fan clocked up to 48-49% of rated output — even with a 47% maximum setting. It’s not clear, however, if GPU-Z is actually reading the fan speed data correctly. Ultimately, it doesn’t matter. What we cared about was whether or not the GPU would throttle under load or maintain its performance. If it’s maintaining its performance at default fan speeds, that’s what matters.
Not all retail cards suffer from this problem
Justin’s Sapphire card, it should be noted, has no problems whatsoever. While I had initially planned to ask him to run before-and-after performance figures using the shipping driver and then the new, AMD-provided driver, it simply wasn’t necessary. At its default fan speeds and stock settings the R9 290 retail sample never throttled at all — and that’s using the same beta driver that caused throttling on our reference card.
In short, this looks like a genuine driver snafu, not a fundamental card problem. We’re going to keep an eye on the issue, because it’s true that the R9 290 is running close to its thermal trip point, but provided end users see consistent performance from the card, our recommendation stands. We’re still hoping to see better third-party coolers in the long run, but the revised driver doesn’t seem to have much impact on the R9 290′s overall noise profile, and it successfully solves the stabilization issue.
Our recommendation stands — though we do wonder how this driver issue made it through QA in the first place…