Laird: DirectX 11 Excitement
Once you’re reading these words, Windows 7 should just about be a retail reality. Leaving aside the unspeakable awfulness of the operating system it replaces, I’m guessing that you’ve probably had your fill of Windows 7 proselytisations. It’s time to take it for a spin for yourself.
However, there is one Win7 feature that’s particularly close to my heart. I reckon I can get away with bringing it up because it’s not actually exclusive to Windows 7. I speak, of course, of DirectX 11. On the face of it, DX11 is just the latest in a long line of multimedia APIs. Great if you have a fetish for fragment shaders, moderately interesting if you’re a keen PC gamer, but otherwise a bit of a niche subject.
DirectX 11, however, is a bit different. There are several reasons to think it will not only be the biggest ever step forward in entertainment tech on the PC, but also properly shake up the way desktop computers process data.
Undoubtedly the biggest change is the introduction of Direct Compute, also known as the compute shader. The idea is to broaden the scope of DirectX – or perhaps more accurately the Direct3D pipeline subset – to general computing rather than just graphics rendering. The only caveat is that a given task must lend itself to parallel processing. In hardware terms, the target component here is the GPU. Despite the recent rise of multicore CPUs, it’s by far the most parallelised PC-compatible computer chip.
It’s also theoretically the most powerful – if only its resources were harnessed for general computing. If that sounds like a familiar refrain it’s because I’ve preached the virtues of using graphics for general processing, otherwise known as GPGPU, on several occasions. It’s a well established concept, complete with several ongoing initiatives, most obviously Nvidia’s CUDA platform.
So far, GPGPU has failed to live up to the hype. I’m convinced DX11 will finally change that. Crucially, it sets out common standards for both hardware and application developers. Unlike with CUDA, you won’t to need to worry about the make and model of your graphics card beyond ensuring it’s the real DX11 deal. Just as importantly, software developers can begin to get their teeth into the challenges posed by coding for GPGPU, safe in the knowledge that there will be a healthy installed base of fully compatible PCs a year or two from now.
Of course, Direct Compute isn’t just about guaranteeing cross-vendor compatibility. It’s also about laying out the minimum hardware requirements to ensure compliant graphics chips actually have the computational chops to handle general-purpose tasks. While previous graphics cards have packed immense theoretical processing power, they’ve also been hobbled by architectures optimised for graphics rendering. Up until the introduction of DX11, for instance, GPUs have only been required to make 256 bytes of internal memory available to each software thread. With Direct Compute in DirectX 11, that balloons to 32kB.
The net result of the combined efforts of Microsoft, Nvidia and AMD should see a huge increase in parallelised application performance. Graphics chips could turn out to be 30 times faster for highly parallised software such as media encoding.
For the record, the first fully DX11-compliant GPU is already on sale. It’s AMD’s new Radeon HD 5870 and it really is a piece of work. Thanks to no less than 2.15 billion transistors, it packs a ludicrous 1,600 stream processors and is claimed to be capable of nearly three teraflops of raw computational heft. To put the latter figure into context, that makes it faster than the world’s fastest supercomputer circa 1999. That was a machine that filled a 230m2 room. Not bad for a single chip.
Direct Compute aside, there are one or two further features that mark DX11 out from its predecessors. For starters, support for multicore PC processors has been much improved. The introduction of hardware geometry tessellation should also make a huge difference to the quality of PC graphics. The sheer number of triangles the new Radeon HD 5870 is capable of processing simply beggars belief. But most significantly, Microsoft has done a much better job with backward compatibility than it did with DX10. You can already upgrade Windows Vista to full DirectX 11 support with a couple of mouse clicks.
Naturally, I would recommend that everyone takes the opportunity presented by the arrival of Windows 7 to go all the way and dump Vista altogether. But I also realise the reality for many is that there are some rather off-putting cost and technical barriers to doing so. It’s nice to know that existing Vista licensees can get all the benefits of DX11 without upgrading to 7. Truly, it’s not often you get something for nothing from Microsoft.
Assuming the architecture is scalable and was designed for large amount of users, you would still need the computing power for that (CPU, memory, DB, network…). And what happens until you reach millions of users and ‘just’ have dozens of thousands of users – how can you stretch your infrastructure to support it?