Mashed Articles

The end is nigh for the modern graphics chip. It genuinely pains me to say that. After all, I’m an unapologetic chip aficionado, someone who loves the technology of integrated circuits for the sake of it. But it’s becoming increasingly apparent that GPUs are over-engineered, increasingly irrelevant and almost definitely not long for this world.

The background here involves a confluence of technological trends. The most ominous of these in terms of the GPU’s longevity as a discrete component is the architectural convergence of CPUs and GPUs. However, one of the most debilitating symptoms of the graphics chip’s terminal malaise is complexity – sheer, pointless complexity. Take Nvidia’s latest uber pixel pumper, the GeForce GTX 480. It weighs in at three billion transistors. That’s getting on for triple the size of Intel’s beefiest PC processor, the six-core Core i7-980X.

If the GTX 480 was any use, that monster transistor count would actually add to the allure. But the harsh truth is that it isn’t – for almost anything. And that makes it dumb. You see, despite the hype regarding running non-graphics applications on GPUs, there’s still very little outside of games that makes more than passing use of a desktop or laptop GPU. More to the point, the number of games demanding a really high-end GPU that are actually worth playing isn’t merely a small number. It’s zero.

Put it all together and you have a terminal mismatch between the cost and complexity of GPUs and their real-world utility. In truth, I’ve felt this way for some time. But it’s the apparent emergence of a radical alternative to established 3D rendering technologies that really brings home how bloated and ludicrous graphics chips have become.

This alleged revolution in rendering comes from a small Australian software startup known as Unlimited Detail. It’s not actually brand spanking new, having been in development for a year or three. But thanks to the random nature of web-based content aggregators, Unlimited Detail was lifted from obscurity recently in a flurry of YouTube-powered publicity.

Anyway, as far as I could tell the basics of this new rendering technology involve ditching polygons in favour of atomic points in 3D space. The claimed result is quite literally unlimited geometric detail. Oh, and the whole thing runs in software at smooth framerates on a conventional PC processor. The GPU doesn’t get a look in until it’s time to spit out the final 2D images.

You hardly need me to point out it all seems too good to be true. So, there was nothing for it other than to go straight to the source and speak to the guys at Unlimited Detail. The technical brains are provided by Bruce Dell, a former supermarket manager, while the business nous comes courtesy of Greg Douglas, a games insider formerly of developers Auran.

The idea of using atoms or points is not new, of course. The really clever bit in UD is the 3D search algorithm developed by Dell. The precise details are UD’s big secret. But according to Dell, “The algorithm takes point cloud data and files it in a certain way so that it can be quickly sorted and accessed.”

When the algorithm searches for points, it doesn’t do so indiscriminately. Instead, it only pulls up a single point for each on-screen pixel being rendered. “We only grab the atoms we need for each pixel, we don’t touch the others,” explains Dell. In other words, the workload depends on screen resolution, not the underlying geometric detail of the scene being rendered. Thus, an impression of unlimited geometry is created.

The UD guys claim the algorithm is so efficient it runs in real-time in a single thread on just one core of a conventional PC processor. Apparently, it will even scale down to simple CPUs in mobile devices.

So far, the only hard evidence for these incredible claims takes the form of a few pre-recorded videos of dubious quality. However, having spoken to the UD pair, I’m happy to confirm they’re not only incredibly passionate, but strike me as completely genuine. It’s potentially extremely exciting stuff.

Still, even if UD works exactly as advertised, the established players in graphics are hardly going to embrace a technology that instantly renders several decades and billions of dollars of investment obsolete overnight. You have to assume Nvidia, and to a lesser extent AMD, will resist the idea strongly. But if Unlimited Detail’s technology gains any traction at all, GPUs really will look sillier than ever.

There’s a delightful story that does the rounds regarding one of the founding fathers of Linux. It’s said that during the early days of the open-
source operating system’s development, this fellow took to attending conferences in complete silence. All attempts to communicate via means other than hand gestures were refused. Instead, he pointed at things.

Apocryphal or not, the tale remains highly relevant today. Our hero’s beef was with the windows-based graphical interface metaphor and its knack for turning us into mouse-pointing morons. Fast-forward a decade or two and astonishingly little has changed. The windows GUI has, you might say, proven to be extremely gluey.

The classic case study is Microsoft’s eponymous Windows OS. Admittedly, early versions of Windows would seem pretty alien to today’s users – but that’s an illusion. Look past the clunky graphics and Windows 95 is largely identical even to Windows 7, Redmond’s latest and greatest OS. Icons, taskbar, the folder metaphor – all are essentially the same as they were 15 years ago.

That’s a long time in any industry, but it’s an absolute eternity in information technology. Along the way, Microsoft has flirted with a few interesting new features. Early betas of Vista included widespread use of virtual folders and the promise of a fully vectorised and hence scalable graphical interface, for instance. But in the end, the retail build of Vista was yet another reskin of Windows NT, just a bit prettier.

Linux and Apple’s Macintosh operating systems have scarcely been any more innovative. More user-friendly and configurable? Perhaps. More polished? Certainly. But both remain firmly rooted in the window-juggling keyboard-and-mouse camp.

Compared to the enormous advances made in computer hardware, it’s all a bit bizarre. Back in 1995, a single-core Pentium processor running at 100MHz or so was your lot. That’s an in-order 3.1 million transistor chip with 8kB of cache memory, for goodness sake. Today, we’re up to six cores, multiple GHz, over a billion transistors and cache pools nigh on double-digits in MB.

If you think that’s merely a matter of scale rather than a new paradigm per se, what about features such as virtualisation or hardware-accelerated 3D graphics? That’s to say nothing of the rapid rise of LCD monitors and more recently solid-state drives. By any sane metric, computer hardware has been in a constant state of revolution. It’s utterly relentless.

So, not to put too fine a point on it, what gives with GUIs? The answer, frankly, is that I don’t know. Over the years, I’ve visited several labs dedicated to advanced interface research, including those of Microsoft and Intel. I’ve even interviewed luminaries from the heyday of interface research, including some who worked at the fabled Xerox PARC lab in Palo Alto. The very people who invented the GUI, in other words. In fact, I reckon I’ve spoken to all the right people. I’ve played with all the latest table-top, touchscreen human-machine interfaces. But I remain essentially clueless. Nothing I’ve seen or heard of is obviously the next big thing.

At this point, Apple’s iPad inevitably hovers into view. A remarkable device in many ways, it’s no good for data input or content creation and therefore doesn’t offer a plausible alternative for desktop computing. However, what it does is underline just how painful the Windows interface is. Once you’ve danced around a few of your favourite websites courtesy of the iPad’s delightfully responsive screen, the scrolly-scrolly, pointy-clicky PC experience seems pretty laughable.

Even a good smartphone can make the PC feel clumsy; I often prefer reading emails on mine. Replying to them is out of the question, but as a viewing device it’s very pleasant and provides temporary relief from what is becoming an overly familiar and oppressive desktop computing experience. You could say the differences are largely arbitrary, but trawling emails on my phone feels like a break from work. That’s got to say something about the tiredness of the windows metaphor.

However you slice it, the PC processor business this year has been all about Intel. First came a revolutionary dual-core chip with a graphics processor shoehorned into the CPU package. At the other end of the scale, a new six-core monster has made Intel’s position at the top of the performance table look even more unassailable. Meanwhile, AMD’s CPU division has released, well, very little. Minor revisions of its dual and quad-core chips has been your lot.

That is about to change. By the time you read these words, AMD will have unleashed its own hexa-core desktop processor, known as Thuban. Based on AMD’s ageing Hammer platform, it won’t be enough to scare Intel’s finest. We’ll have to wait until 2011 and the arrival of AMD’s new Bulldozer CPU architecture before there’s any chance of that happening, but Thuban should make for an interesting addition in the meaty mid-range of the CPU market.

Of course, I’ve chewed the fat regarding the prospects for Bulldozer in PC Pluses passim. But it’s actually the chip scheduled to arrive between Thuban and Bulldozer that should mark the beginning of AMD’s renaissance. Codenamed Llano and due late this year or early next, it’s AMD’s first stab at a CPU-GPU ‘fusion’ processor. Intel may have beaten AMD to market with such a device, but it’s looking like AMD might have the edge.

For starters, AMD has much better graphics technology than Intel. That’s unlikely to change any time soon, given the recent cancellation of Larrabee, Intel’s stillborn effort to engineer a stand-alone graphics processor. Indeed, thanks to some recent disclosures from AMD, the full implications of its superiority in graphics are becoming clearer. Images of the Llano die have been circulating and reveal that the graphics core is big. Really big. It consumes nearly half the available die space. Remember, this is a 32nm quad-core processor. That means AMD has invested lots of transistors in the graphics.

Extrapolating from such images is always perilous, but I’m confident it will be by far the most powerful integrated graphics chip ever. To put it into context, the fastest integrated graphics core AMD currently makes has 40 stream processors. Llano is thought to have as many as 480, making it over 10 times as powerful. Indeed, the Llano graphics core looks more powerful than AMD’s fastest discrete desktop GPU of just a couple of years ago, the Radeon HD 2900.

Comparisons with Intel’s current Westmere fusion chips are tricky, not least because they are dual-die constructions combining a 32nm dual-core processor with a 45nm graphics chip in a single package. Suffice to say that in terms of rendering grunt, Westmere doesn’t even rank as the most powerful current integrated GPU, much less threaten Llano. Of course, by the time Llano appears Intel will be tooling up for its upcoming Sandy Bridge family of chips. Like Llano, Sandy Bridge derivatives will be quad-core, single-die chips. Intriguingly, die shots reveal the graphics core in Sandy Bridge represents at most one fifth of its total area.

All of which means we have two very different looking approaches from AMD and Intel when it comes to fusion processors. You could say each plays to its strengths. AMD has great graphics and relatively weak CPUs, so stuffing a big GPU into its fusion processor allows for a strong narrative. Want a great visual experience from your low-end PC? Then you need AMD. Meanwhile, Intel will be emphasising the superiority of its traditional CPUs, something it will have no problem demonstrating.

On the desktop, it looks like a win for Intel. Anybody who wants serious graphics power will simply drop in a discrete video card (though there’s a good chance that video card will come from AMD). It’s a different story for laptop PCs, however. Power use is the driving factor here, and that means a heavyweight discrete GPU is a non-starter. Likewise, outright CPU performance is less of an issue for mobile systems.

What AMD will therefore be able to offer is a more balanced package. Its CPU won’t be the fastest, but it will be good enough. Meanwhile, it will have by far the best integrated graphics core, all the better for everything from gaming to high definition video playback. What’s more, if general processing on the GPU ever takes off, well, that will just be gravy. All this, of course, is before the arrival of AMD’s new Bulldozer CPU architecture. If that’s any good, Intel really will be under pressure.