AMD RX 480 review: The best budget graphics card—but for how long?

Brave? Foolhardy? Desperate? Whatever you might think about AMD's decision to cede the top end of the graphics card market (at least for now) to Nvidia and launch the mainstream-focused RX 480 instead, the fact remains that for £180/$200 it's the best graphics card you can buy. It's faster than Nvidia's GTX 970 and (mostly) faster than an R9 390, making it more than powerful enough to meet the minimum spec for virtual reality—and it'll blitz through demanding 1080p games at a smooth 60FPS, too. It even does a decent job at 1440p, so long as you're fine with dialling down a few settings.

As a consumer product, then, the RX 480 is a success, even if one of AMD's core pitches—that it'll help drive VR adoption—is a little suspect. VR headsets still cost well over £500, after all.

But—and sadly, there always seems to be with AMD—the RX 480 is not a great debut for Polaris 10, its first GPU based on an all new, theoretically more efficient 14nm FinFET manufacturing process. At 150W, the RX 480 sits in the same power envelope as the GTX 1070 yet offers less performance. It runs hotter, too, hitting 80 degrees Celsius, even struggling to hit its advertised boost clock at times—and that's in a big, well ventilated case. Compared to AMD's previous cards, it's an improvement, but those were always power-hungry beasts, and the bar has since been raised.

None of that matters to consumers hell bent on getting the best bang for their buck, of course. And with the sub-£200/$200 market being the most popular for discrete graphics cards, the RX 480 is a smart play to bring AMD's market share up to more respectable levels. But Polaris 10 is already being pushed hard in the RX 480, and graphics nerds hoping that AMD would regain its performance lead in the future might wonder whether AMD set out to make a mainstream card or whether—thanks to less than stellar efficiency improvements with Polaris—its hand was forced.

Polaris 11: The new old

Indeed, a quick look inside the RX 480 hints at why performance hasn't taken quite as big a leap as you would imagine. There's a new chip, dubbed Polaris 10, which... well, just isn't very new. It's very similar to the older GCN 1.2 architecture used by the Fiji GPUs that powered the Fury X and the R9 290X before that, albeit on a smaller scale. There's a single main Graphics Command Processor, four Asynchronous Compute Engines (ACE), two hardware schedulers, and 2MB of L2 cache sat atop the RX 480's 4GB or 8GB of memory. In total, there are 36 compute units (CUs) and 144 texture units pumping out just over five teraflops of FP32 performance.

Specs at a glance

That brings the RX 480 in line with, or above the likes of, the GTX 970 and R9 380 and just shy of the 5.9 teraflops of the R9 390. Take the GTX 970 out of the picture—which is priced on Nvidia's more premium scale—and despite all the pre-release hubbub, the RX 480 simply slots between two of the company's existing cards. A groundbreaking product this is not.

Still, there are a few features in the RX 480 that you won't find in older graphics cards. There's DisplayPort 1.3 HBR, HDMI 2.0b (at last!), and support for HDR content. Without any consumer HDR PC monitors around, you'll have to pick up a TV in order to actually view HDR content, although AMD says it's working with monitor manufacturers to get models on the market as soon as possible. Native FP16 (half-precision computation) support has been introduced for the first time, too, and while that will mean little to gamers, those into heavy compute tasks such as computer vision or machine learning stand to benefit from the reduced memory use of FP16 instructions.

On the outside, the RX 480's reference cooler—which many of the launch cards are likely to be based on—takes its cues from the Fury X. The soft-touch, dimpled plastic top panel looks slick without being ostentatious, while the simple blower design does a good job of shovelling hot air away from the GPU, even if it is a tad noisy under heavy load. It's an extended design, too, with the fan hanging over the edge of the PCB. Expect smaller partner cards with custom coolers ideal for teeny ITX builds to appear in the near future.

One thing to note about the RX 480 is its memory. AMD has talked up the RX 480's fast 8GB/s GDDR5 memory since its announcement, which—when tied to its 256-bit bus—results in 256GB/s of memory bandwidth. However, it turns out that only the more expensive 8GB version of the card (£215/$230) will feature 8GHz GDDR5, with the cheaper model featuring slower 7GHz GDDR5 for only 224GB/s of bandwidth. Even then, AMD isn't promising that all 8GB versions will feature the faster memory, saying that "you may see configurations that vary from our reference spec."

The difference between the two isn't huge, and there's only likely to be a few frames in it in real-world performance. But since the 8GB, 8GHz GDDR5 RX 480 is what's tested here—and indeed what AMD is sending out to all reviewers at launch—double check exactly what speed the memory is running at on the card you're looking to buy before parting with any cash.

But there are some architectural improvements

While Polaris 10 features the same high-level architecture as older GCN 1.2 GPUs, there have been some low-level changes that AMD says have given Polaris a 15-percent boost in performance per compute unit over previous generation architectures. These improvements are similar to those AMD made in its Bristol Ridge APU line, which also featured a near-identical high-level architecture to its predecessor.

First up is a new power calibration routine that takes place at boot time, which removes the safety limitations associated with variations in consumer power supplies versus those used in manufacturing. By comparing the voltage analysis run during manufacturing with the boot time readings, the voltage regulators on the board can be precisely dialled in to match. This eliminates the waste power associated with accommodating individual motherboards and power supplies, allowing for more stable boost clocks.

There's a new reliability tracker that tracks how much the chip's transistors degrade due to heat and high voltages, which can cause the chip to slow down as it ages. Again, this allows AMD to remove the limitations imposed on every chip in order to safeguard its reliability. The tracker monitors instances where heat and voltage are within safe limits, allowing the chip to boost at higher clock speeds for longer duration.

Much improved power gating, a new multi-bit flip-flop (say that fast three times), and a larger L2 cache also help to bring GCN's efficiency up. That AMD's engineers have extracted yet more power efficiency, however small, out of an architecture that was already stretched to its limits in the Fury range is an impressive feat. Although, it's questionable how much longer AMD can keep it up; sooner or later GCN will need a massive overhaul to stay relevant (fingers crossed for Vega!).

You WattMan?

Launching alongside the RX 480 (and only compatible with the RX series) is a new version of AMD's Overdrive overclocking tool, which is now (amusingly) dubbed WattMan. I've always been a fan of AMD Overdrive, since it made overclocking so simple and didn't require downloading an external tool and running it in the background like with Nvidia cards. While WattMan remains built into AMD's driver software—it's found under the Global Settings page in Games—the simplicity has been well and truly ditched in favour of extreme overclocking control.

WattMan presents you with an intimidating array of sliders, numbers, and rotary tools that allow you to adjust the GPU frequency per voltage state. The idea is similar to that of Nvidia's GPU Boost 3.0, in that every GPU produced by the fab has slightly different characteristics and thus can attain different clock speeds at different voltages. A traditional overclock doesn't take those characteristics into account and instead applies a linear frequency increase across the board. In theory, by creating a more nuanced frequency curve, you can extract more performance from the GPU before running into crashes.

It's a neat idea, although—like most overclocking tools—the UI could use some work. I generally avoid review guides that are handed out to press for anything other than fact-checking (consumers don't have them after all), but after half an hour of fiddling around with the various knobs and sliders in WattMan to no effect, I gave in and decided to read the manual. Even with the solution, enabling overclocking isn't a particularly intuitive process. The important buttons are small and hard to find, while the layout demands that you stretch it out across an entire display (tested at 1440p) in order to read everything clearly.

The process, it turns out, involves flicking the switch under the temperature section at the bottom to change it from automatic to manual. That unlocks the power limit slider underneath, which you can then ramp up in order to adjust clock speed. From there, you have to move up to the GPU section where you can raise the frequency percentage slider to your desired overclock. AMD says this simple overclock should be enabled by default, but I had to click the frequency percentage toggle switch to enable it. Those that want to mess around with individual voltage point overclocking can just drag the vertical sliders.

Having such deep control over overclocking is no bad thing, but the addition of a simple mode that removes the extraneous fluff would do wonders for improving usability.

Fan speeds and temperature targets have been given an overhaul, too, with much deeper options including the ability to adjust the maximum temperature before system clocks are reduced to cool down the GPU and adjust the target temperature before the fan speed is raised to cool down the GPU. There's even a data capture mode that allows you to record the peak and average GPU activity, temperature, fan, and GPU and memory clocks for up to 20 minutes while running an application. For power users wanting to know how hard a game is hammering the GPU the hardest, it's an interesting tool.

Having said all that—and having to deal with the shame of reading the manual—for some reason my sample of the RX 480 refused to overclock at all. Whether it was a bad sample, flaky drivers and software, or simply that there wasn't any headroom for overclocking, no matter what I did, the PC would crash. We're investigating this issue with AMD, but for now there aren't any overclocked results in the benchmarks below.

Performance

The RX 480 was tested with a suite of games on the Ars Technica UK standard test rig, including three games that use DirectX 12. There's still no reliable way to capture frame data for DX12 games without a dedicated hardware setup just yet, but for everything else there's a 99th percentile score, which shows the minimum frame rate you can expect to see 99 percent of the time. This is a great way to highlight the comparative smoothness of games—the higher the gap between the average of the 99th percentile, the more jittery a game feels.

Each game was tested at 1080p, 1440p, and UHD (4K) resolutions at high or ultra settings at stock speeds. Of note, however, is that while the RX 480 claims to feature a core clock of 1120MHz and a boost clock of 1266MHz, during testing it would often top out at 1206MHz. Manually adjusting the fan speed up slightly rectified the problem, but that the RX 480 struggled to hit its advertised boost clock in a big, well-ventilated case is slightly worrying.

On the synthetics and science side there's the standard 3DMark Firestrike benchmark (again, run across three resolutions), as well as LuxMark 3.0, CompuBench, and FAHBench (the official Folding@Home benchmark) to test compute performance.

(A special shout out goes to Overclockers UK, which kindly loaned Ars UK an XFX R9 390 and MSI GTX 970 for the purposes of this review. Suffice it to say, I hear your cries about getting more relevant graphics cards on the charts.)

The good news first: as AMD promised, the RX 480 is indeed ready for VR. It's as low as four percent faster than a GTX 970 (the minimum recommended VR spec) in Metro: Last Light and as high as 25 percent faster in Hitman. Sure, the on-the-fly image quality adjustments of VR games might not ramp up very high, but if it's the price of graphics cards that's putting you off VR, the RX 480 is the solution.

At 1080p, it's good, too, running nearly every game on maximum settings at or above 60FPS. Minimum frame rates were good, too, particularly when compared to the R9 290X. Both cards turn in similar average frame rates, but there's a big boost in minimums for the RX 480—those stuttering and frame pacing issues that plagued older AMD cards seem to have finally been put to rest.

The bad news is that compared to the more expensive R9 390, the RX 480's scores are less clear cut. In more modern games like Hitman, Rise of the Tomb Raider, and The Division, it's faster at 1080p by around five percent. However, the situation is reversed in older games like Crysis 3, Bioshock Infinite, and Metro: Last Light where it's around five percent slower. The RX 480 fares particularly badly in Battlefield 4, where it's slower than even a GTX 970, although I suspect that's a driver issue more than anything else.

The RX 480 passes for a good 1440p card, too, provided you can live with a locked 30FPS or dialling back a few settings in order to boost the frame rate. At the higher resolution, however, the R9 390 begins to claw back its lead, with its higher memory bandwidth and greater number of CUs overcoming the RX 480's higher clock speed. At 4K, the difference is far more pronounced, but outside of simple games I wouldn't recommend either for 4K: the performance just isn't there to make even smooth 30FPS gaming feasible.

Good, not great

So, here we are, after months of speculation and expectations over AMD's first foray into FinFET with its 14nm Polaris GPU. What the company has produced might not be what everyone wished for (who doesn't want an Nvidia-busting high-end card?), but there's no denying that the RX 480 is a solid graphics card that lets you access VR gaming at a previously unheard of price. And if you're not into VR, it'll churn through pretty much any 1080p game you can throw at it, too.

And yet, this sort of performance for £180/$200 is to be expected. If the RX 480 was significantly faster than an R9 390, perhaps even challenging a 390X or a GTX 980, it would be a fantastic graphics card. With greater efficiency, it might even be a groundbreaking one. Instead, the RX 480 maintains the status quo, filling a gap in AMD's product stack without disrupting it. Only the R9 380 and 380X are rendered redundant, and even then only for new buyers.