AMD has released an updated driver for its RX 480 GPUs that should permanently resolve the GPU’s power draw issue. For those of you just tuning in, the brand-new Radeon RX 480 launched last week with an unusual power configuration. Specifically, it drew more current from the PCI Express slot on the motherboard than is allowed by the official PCI-SIG specification. While the chances of a problem or failure were low, AMD promised an updated driver that would resolve these issues permanently. The company released Catalyst 16.7.1 yesterday afternoon, and we’ve put the driver through its paces.
There are two changes being introduced in the 16.7.1 driver (the launch driver for the RX 480 was Catalyst 16.6.2 for those of you keeping score at home). First, AMD has changed how the GPU draws power from the system. Instead of splitting the load roughly 50/50 between the PCI Express slot and the six-pin power connector, the card now draws more power from the six-pin slot and less from the PCI Express slot. This change is baked into the driver and automatic. Testing from sites like PC Perspective and THG has confirmed that this shifts roughly 10W of power over to the PCI Express six-pin connector and reduces the amount of current being pulled off the PCI Express slot. While the six-pin connector is now operating somewhat out of spec, it’s also designed to handle 8-9 amps per pin and is overengineered by a considerable margin.
Tom’s Hardware’s testing shows that the amount of current pulled from the mainboard has fallen from 6.7 amps to 5.9 amps. While this is much closer to the 5.5 amp specification defined by the PCI-SIG, it doesn’t pull the GPU 100% into compliance. That’s where the driver’s other new feature comes in: A new (badly named) “Compatibility Mode” that throttles the GPU back slightly to ensure that the chip never exceeds PCI-SIG specifications in any particular. Compatibility mode is off by default in the Catalyst 16.7.2 driver; end users who are concerned about power draw can enable the feature manually if they choose.
Unlike the baseline fix, which has no impact on performance, activating compatibility mode does have a small impact on how the GPU performs. AMD characterized the impact as roughly 3%, but we re-ran our own suite of benchmarks to measure the drop. Some of you were concerned that this change would significantly impact RX 480 performance — but that’s not what we saw.
We’ve gone back and included additional decimal places to confirm that the changes (or lack thereof) aren’t being hidden by rounding. Compatibility mode performance is on top, followed by Catalyst 16.7.1 without compatibility mode enabled, followed by the 16.6.2 launch driver.
As you can see, the degree of difference across the three drivers is extremely small. The RX 480 is slightly faster in Total War: Rome 2, tied in Ashes of the Singularity, tied in Dragon Age Inquisition, slightly faster (5.5%) in Shadow of Mordor, roughly 5% slower in compatibility mode when playing COH 2, very slightly slower in BioShock Infinite (less than 2%), and the same speed, for all intents and purposes, in Metro Last Light Redux. Overall, that works out to some small gains without Compatibility Mode enabled and flat performance with it. Frame timing and smoothness are also unaffected.
Finally, we tested total power consumption at the wall. While we can’t measure the individual power rails, we can check to see how the GPU is behaving as a whole.
In other words: AMD appears to have delivered precisely what it said it would. Total GPU power consumption drops very slightly with the compatibility mode enabled and increases very slightly with the new driver otherwise, but doesn’t shift significantly in either direction.
There have been reports that undervolting the RX 480 can improve its performance, so I took the GPU out and tested it to see what I could find. In theory, the adaptive voltage and frequency scaling (AVFS) that AMD uses for the RX 480 should have captured most of the headroom available to the chip before — and that seems to be the case.
In his article at LegitReviews, Nathan Kirsch discusses how dropping the voltage on his RX 480 from 1100mV and 1137mV at the highest power states improved performance by roughly 3%. Our sample GPU uses tighter voltages — just 1068mV at State 6 and 1106mV at State 7. While these differences might seem too small to matter, power consumption in GPUs rises roughly as the cube of the voltage — meaning small changes can yield significant differences. In our case, undervolting the GPU to 1050mV and 1090mV only improved performance by less than 2% (from 47.33 to 48.2 FPS). Giving the board an extra 50% power headroom didn’t make a difference in overall performance.
While our chip’s voltage may be lower than some of the cards on the market, that doesn’t seem to translate into better overclocking headroom. Our chip wasn’t stable much above 1300MHz, with a maximum frame rate of 49 FPS (up from 47.33 FPS) in Metro Last Light Redux. We’re not really comfortable pushing the RX 480 faster at this point, mostly because it’s not clear if doing so will result in pulling more current from the PCI Express slot than it is specced to handle. AIB cards from AMD’s partner vendors with more robust power circuitry could unlock more headroom, but for now it seems wise not to push this particular envelope — especially when some users are seeing equivalent performance gains and better power consumption just from undervolting the card. We’ve asked AMD for additional information on how the RX 480 will handle overclocking power distribution and will update this if we hear back.
We’re glad to see AMD responding to this situation quickly and decisively, with a driver update that solves the problem without impacting overall performance. The 16.7.1 driver doesn’t change any of our conclusions about the RX 480 — this GPU is still vastly more power efficient than any previous $200 card from AMD, and it offers a great combination of performance and features at its price point.
But — and I really can’t stress this enough — this is now the third major launch in a row that AMD has self-sabotaged through a series of completely unforced errors. Back in 2013, the R9 290 and 290X were tarnished by poorly designed, noisy reference coolers that struggled to keep both GPUs below 95C. Later designs from AMD’s board partners proved that the R9 290 and 290X could be cooled more effectively and quietly, but too late — people looked at the reviews of the reference models and thought the GPUs were cacophonous furnaces.
The R9 Fury X last year may not have been noisy, but it still had a noise problem, thanks to bits of glue that caused a particularly annoying high-pitched whining unrelated to the usual problem of coil whine on high-end cards. Again, AMD fixed the issue relatively quickly, but it still fed perceptions that the company wasn’t firing on all thrusters.
Last year, AMD reorganized its entire graphics division under Raja Koduri, launche and set out to rebrand its new RTG (Radeon Technology Group) as distinct and different from what had come before. Yet here we are again.
Problems like this don’t just feed negative perceptions of AMD. Worse, they obscure the very real improvements the company has made to its driver software, its driver cycle release time, frame pacing, and underlying hardware. When third-party vendors release their own custom RX 480 designs, there’s going to be an ongoing discussion about whether or not their cards avoid this problem and, if so, how they avoid it. A relatively minor flaw will end up embedded in future reviews because someone signed off on a configuration that shouldn’t have shipped, potentially compromising how consumer’s view the product in the process.
2016 isn’t over. With Vega launching later this year, RTG still has an opportunity to demonstrate that it can build and launch an uncompromisingly great GPU with no errata or issues that require significant post-launch testing to address. Here’s hoping they do.