RX 480 reference card, courtesy AMD.
The rumor mill has yet to cease churning with word of AMD’s upcoming RX 590 graphics card, based on GlobalFoundries’ 12LP process. The RX 590 is said to utilize a respin of the aging Polaris chip, known as Polaris 30, shrunk down to the 12nm node from 14nm. This may give AMD another ~200MHz of headroom to work with, but is it enough to make a dent in NVIDIA’s share of the market?
For background, AMD initially launched the Polaris architecture with the RX 480, using the 14nm Polaris 10 GPU in 2016. An optimization, known as Polaris 20 and released as the RX 580, was released in 2017, providing slightly higher clock speed headroom on the 14nm process at the expense of power consumption. Polaris 30 marks the third refresh of the Polaris architecture for AMD, two years later, while NVIDIA has already moved on from Pascal to Turing. However, Turing is currently limited to the ultra-high end (>$500) market. As a result, the RX 590 will be going up against the same GTX 1060 that the RX 480 battled two years ago, and that the RX 580 is still in a dead heat against. The specifications of these three cards are not substantially different:
|Graphics Card||Radeon RX 480||Radeon RX 580||Radeon RX 590 (TBC)|
|GPU||14nm Polaris 10 XT||14nm Polaris 20 XT||12nm Polaris 30 XT|
|Core Config||2304 SP, 144 TMU, 32 ROP||2304 SP, 144 TMU, 32 ROP||2304 SP, 144 TMU, 32 ROP|
|Memory||8GB 256-bit GDDR5-8000MHz||8GB 256-bit GDDR5-8000MHz||8GB 256-bit GDDR5-8000MHz|
|Thermal Design Power||150W||185W||185W?|
The rumored 15% clock bump, given linear scaling, would put the RX 590 decidedly ahead of the RX 480/580 and GTX 1060, but still closer to 1060 levels of performance than 1070 (much less 2070). But can we expect linear scaling?
The main issue I see with Polaris 30 is that, according to rumors, it’ll be using the same memory controller and the same 8Gbps GDDR5 as the previous Polaris cards. The problem is that Polaris is moreso limited by memory bandwidth than it is by raw shading, texturing, or rasterization performance. To a certain point, depending on game/workload, overclocking the memory is more beneficial than overclocking the core. AMD’s equally-performing card from the previous generation, the R9 390X (Hawaii), utilized a 512-bit bus with 6Gbps GDDR5, delivering 50% greater bandwidth than Polaris. More efficient compression algorithms (36% more, to be precise, not 50%) and other optimizations led to this bandwidth deficiency being negligible at the original 1266MHz stock clock, but how far can AMD push the envelope before it becomes pointless?
And moreover, it being two and a half years since Polaris launched, how did AMD lack the foresight to anticipate this refresh and the need for faster memory? 8Gbps may be the limit for stock GDDR5, but NVIDIA (or their board partners) utilized factory-overclocked 9Gbps GDDR5 for certain 1060 models. Given that AMD will presumably be launching a cut-down variant of this GPU, it would make sense for them to use binned chips to deliver higher bandwidth on the 590, then use the low bins for the cut-down card. However, according to current rumors, this will not be the case.
Another option AMD could have gone with would be to redesign, if nothing else (they didn’t redesign anything at all apparently), the memory controller. A 384-bit memory controller could provide for 384GB/s of bandwidth at 8Gbps, the same offered by a stock R9 390X. This seems a bit excessive for Polaris, so they could instead use 7Gbps GDDR5 and yield 336GB/s, which is more than enough, and offset the added cost and power consumption of the larger memory controller. Normally, this would also mean having to increase the rasterizer (ROP) count to 48, though if this were cost prohibitive AMD could have stuck with 32 on 384-bit as they did with Tahiti. A 384-bit, 48 ROP Polaris at 1600MHz though? Would that not be a 1070 competitor?
Practically coinciding with the launch of Polaris, NVIDIA launched Pascal with GDDR5X, which provided a bump to a ~10-11Gbps out-of-the-box data rate. Turing, launched this summer, uses GDDR6 running at 14Gbps. Across a 256-bit bus, 10Gbps delivers 320GB/s, and 14Gbps delivers 448GB/s — the same as the GTX 1080 and RTX 2080, respectively. If AMD could have simply redesigned the memory controller with the node shrink, instead of absolutely nothing at all, even 10Gbps GDDR5X would alleviate the bandwidth bottleneck, delivering a 25% increase in bandwidth versus 15% for the core clock. The main problem here is that memory is expensive, and neither of the newer memory technologies are being produced in particularly large quantities.
This is not to say that the RX 590 will be a particularly bad card for the price when it launches. It’s expected to perform perhaps 10% better than the GTX 1060 at approximately the same retail price. But the 1060 is a 2 year old card, consumes far less electricity, is about to be refreshed with GDDR5X memory itself, and is likely to be replaced by a 2060 in a few months. AMD shouldn’t be refreshing Polaris a second time to edge past it, they should have an all-new chip that decidedly beats it after all this time.
Ultimately, what it comes down to is that AMD is designing the RX 590 as cheaply as they possibly can. Their R&D budget is evidently minimal. From what we’ve seen so far, this card will not deliver a single change except moving to the 12nm node and taking the ~10-15% extra frequency that comes along with it. If it’s similar to the 12nm shrink they did for Zen, they won’t even increase the density of the design, they’ll just increase the space between die elements to improve heat dissipation and frequency potential. After two years, this is the best they can do, finally beating the 1060 when NVIDIA is already starting to roll out 2000-series graphics cards. Now that their GPU division has gotten an overhaul with the departure of Raja Koduri, it’s about time their GPU architecture gets one too — they need it, fast.