At AMD’s Radeon Technologies Group (RTG) tech summit last month, corporate vice president, product CTO, and corporate fellow Joe Macri described fabrication process advances as being “like Christmas” for the company’s engineers. The last time the company took the wrapping paper off a graphics card benefiting from a process tech advance was late 2011’s Radeon HD 7970 and its 28-nm Tahiti GPU. Since then, GPUs from both AMD and Nvidia have been fabricated on 28-nm process technology.
That long period of process stagnation is coming to an end. Around the middle of this year, AMD will release GPUs built with a new architecture it’s calling Polaris, and those chips will be made using FinFETs, a type of 3D transistor. FinFETs bring some tantalizing improvements for performance and power efficiency to the table. Judging by the company’s excitement over Polaris, FinFETs are just the kind of Christmas present the company’s engineers have waited so long for.
Before we talk about FinFETs and their implications for Polaris’ performance, though, let’s examine some of the guiding lights that AMD is using to design its future products.
Charting the course
AMD’s Raja Koduri, senior vice president and chief architect for the RTG, says the division is working on graphics products that can deliver both “fast pixels” and “deep pixels.” The fast-pixel problem comes from the demand for graphics cards to deliver more pixels at ever-greater speeds, a challenge made more difficult by the demands of present and future VR hardware.
For example, Koduri says twin 4K displays in a head-mounted device will require addressing 16.6 megapixels per refresh. If that figure isn’t intimidating enough, the company would eventually like to deliver enough pixels to drive a VR headset with twin 16K displays for a “truly immersive VR experience.”
For that to happen, Koduri says we’ll need an increase in graphics processing performance of at least a thousand times that of today’s chips. Furthermore, he believes that photo-realistic rendering for such a device would require graphics processors to deliver over one million times the performance of the hardware we have now. Koduri estimates that we’d need over 20 years of performance increases from Moore’s Law alone to reach that point, and he isn’t content with waiting that long.
The RTG believes that bridging this enormous gap is going to require new, more efficient rendering approaches born from techniques like light-field rendering, texture-space rendering, and foveated rendering in tandem with more powerful hardware. Koduri hopes that AMD’s GPUOpen initiative, intended to foster more collaboration and information-sharing among game developers, will help those devs to begin cracking some of the challenges of that grand goal on the software side of the equation.
We’ve already seen some of the company’s plans for “deep pixels” in its plans for FreeSync and ultra-high-definition content on Radeon graphics cards this year, and Polaris chips are one way the company will begin to deliver on those promises.
We aren’t going to be diving deep into the microarchitecture of the Polaris GPU today, but we do know a few things about it. Polaris includes a number of fourth-generation GCN graphics cores with improvements for primitive discard accelerators, hardware schedulers, and instruction pre-fetch. AMD also says it’s improved overall shader efficiency in fourth-gen GCN, and Polaris chips will get some form of memory compression, too. Polaris’ display block will support features like HDMI 2.0a and DisplayPort 1.3, and its multimedia features will include H.265 Main 10 decode at resolutions up to 4K, and 4K H.265 encoding at up to 60 FPS.
AMD says Polaris chips have been designed specifically for fabrication with FinFETs. Unlike planar transistors that are laid down in flat layers, a FinFET is built by “wrapping” the transistor’s gate around a three-dimensional fin of silicon. For some more information about the general principles behind FinFETs, check out this Intel presentation (PDF) and video.
Normally, we’d describe those transistors using a feature size like 28 nm or 14 nm, but AMD didn’t share details of the exact processes it’s targeting for the production of Polaris chips at its summit. Even if we don’t know the process that AMD will use to produce Polaris chips right now, FinFETs have a lot to recommend them. The increased surface area of the FinFET’s gate-channel interface offers much more control over the transistor. AMD says this design means a FinFET transistor can switch on and off faster, carry more current when it’s operating, and consume less power when it’s not.
FinFETs have other useful characteristics that make them appealing versus planar transistors. Joe Macri told us that FinFETs are easier to build, easier to characterize, and more uniform than their 2D cousins. That uniformity is important. Macri noted that the performance of a chip is set by its slowest device, and power usage is set by its leakiest. FinFETs exhibit higher overall performance and lower overall leakage compared to 28-nm planar silicon, as illustrated by the hypothetical distributions of the “clouds” above.
All told, AMD promises that graphics chips built with FinFETs will offer a big performance-per-watt increase. The company says it can keep the performance of a given FinFET chip the same versus a 28-nm planar chip while reducing its power consumption by up to 50%-60%. That performance-per-watt improvement means Polaris GPUs could provide what AMD calls “better-than-console” performance in devices like thin-and-light gaming notebooks and small-form-factor desktops.
The company did show working Polaris silicon at the RTG event, and there does appear to be something to its performance-per-watt claims. Although we weren’t permitted to see any graphics cards with Polaris chips up close, the company brought out a pair of small-form-factor systems built with Intel motherboards and Core i7-4790K CPUs to show what one form of its new chip can do.
Those PCs appeared to be inside tiny Cooler Master Elite 110 cases. One held a Polaris graphics card, while the other used an Nvidia GeForce GTX 950. The Polaris PC used about 90W to run Star Wars Battlefront at 1080p with a 60-FPS cap, while the Nvidia-powered system needed about 153W for the same load.
We wouldn’t apply or recommend a 60-FPS cap in games that don’t require one, so it’d be interesting to see how AMD’s fight would play out without that restriction in place. Even so, that small peek at Polaris’ performance has us eager to dig deep when those next-gen Radeons arrive later this year. That day can’t come soon enough.