AMD announces the new Polaris architecture

More than a name change but that is included

AMD Radeon Logo 2013AMD is undergoing a name changing revolutions with their new 4th Gen GNC based Polaris architecture. For those not up to speed with Crimson, under the rain of Raja Raja, almost everything GPU at AMD is going through a conceptual and nomenclatural revolution.

What was Catalyst is now Crimson, what was once Arctic Islands is now Polaris, and so on and so forth. This may sound like mere window dressing but it really isn’t, starting with what was once Greenland, the changes are quite real but there is only so much you can do to hardware in the time frames involved. That said AMD is going to make a big step forward.

If you have been watching the GPU world for the past few years, you might have noticed that AMD has not launched any new architectures for a long time. GPU-years are a higher multiple than dog-years, that is the long way of saying their don’t age well. Neither do CPU cores and they suffer from the same malaise as AMD’s GPU cores, what SemiAccurate calls the Seifert effect. Remember when an unnamed AMD interim CEO paused development as a way of saving money. What was the delight of Wall Street at the time lead to the loss of a lot of top talent, they saw this bubble coming but luckily that will soon be history.

With that AMD is coming out with its first real new architecture in many many many years, Hawaii and Fury were just retreads with some low hanging fruit added here and there, not real changes. This showed up in the power figures, or at least the performance per Watt figures, and a few other areas. Polaris changes all of that with one demo, Star Wars Battlefront running at 1080p.

This barely interactive movie that some pretend is a game was run on two systems, one with an unspecified Polaris GPU, either Baffin of Ellesmere, not the big one, and a ‘competitive’ GPU. Both were the effective minimum you could use to run Battlefront at 1080p smoothly, and both were run on otherwise identical Intel CPUs.

The net result was the AMD GPU pulled 88W (Note: 86W is the official number and we saw numbers above and below that figure, 88W is the number we wrote down), and about 150W for the competitor (Official number from AMD – 140W. Same disclaimer from us here too). Polaris finally brings AMD back to modern performance per Watt ratios. The down side to this is if about half the gain is due to process, Nvidia will be there too, just far later.

So how does AMD get to those figures, and what does Polaris bring to the table? The big bang is from the process, it is being built on an officially unspecified 14/16nm process. The chip we held had Globalfoundries markings on it though, take that hint as you will but officially it could come from anywhere. This chip is aimed th 60FPS gaming without cripplingly low settings in a thin and light laptop, a mark AMD should hit with ease. More interesting is what the APU version looks like, plus they won’t have as abjectly broken 3D drivers as Intel. It could be tempting, very tempting.

AMD Polaris architecture block diagram

This is as detailed as you get for now

The diagram above is about as specific as AMD gets for architecture at the moment, IE not very. For more you will have to wait until Q2 or “mid-2016” officially for more detailed info which AMD promised to actually give out this time. The list of official changes for 4th Generation GCN are a primitive discard accelerator, a new hardware scheduler, instruction pre-fetch, improved shader efficiency, and memory compression.

Most of these are newer and better but the memory compressions is a lot more important than it sounds. Even if exact details are not available, the major constraint on modern GPUs is memory bandwidth, and it is also a massive power hog. Fury solved this fairly elegantly with HBM but that is expensive and for the moment, not suited to lower end devices. This tech will migrate down in a generation or two, we are watching the supply chain closely, but for the moment it is a pain point. If you can compress memory by only 10% you just added 10% more performance or saved 10% more I/O power. If you got better numbers, and we think AMD will, it is that much more beneficial. Of all the core changes, memory compression is likely the most important one.

That isn’t to say the uncore is unchanged, there are quite massive changes there too. It all starts with a new display engine capable of HDMI 2.0 and DP1.3, basically Polaris can do 4K natively and without compromise over a single cable. To use the industry vernacular, about $&*^%ing time, even if AMD is first to this late mark. Added to this is a new multimedia engine that can do h.265 Main10 decode at 4K and encode at 4K60. Although it wasn’t explicitly stated, it is likely that the new GPUs will support HDR outputs and video decode as well, the extent of this will have to wait for the architectural disclosures later in the year though.

In a nutshell that is what AMD brings to the table. It starts with a philosophy change that is first manifested in a naming change but it won’t stop there. The Seifert effect has been overcome and hopefully the results will look like the post-Coyote and Roadrunner era, the early signs are promising. The hardware is finally thoroughly modern as far as power efficiency, process, media standards support, and outputs. Q2 can’t come soon enough.S|A

The following two tabs change content below.

Charlie Demerjian

Roving engine of chaos and snide remarks at SemiAccurate
Charlie Demerjian is the founder of Stone Arch Networking Services and is a technology news site; addressing hardware design, software selection, customization, securing and maintenance, with over one million views per month. He is a technologist and analyst specializing in semiconductors, system and network architecture. As head writer of, he regularly advises writers, analysts, and industry executives on technical matters and long lead industry trends. Charlie is also available through Guidepoint and Mosaic. FullyAccurate