Your next PC

If we had told you in 1995 that graphics accelerators in ten years time would be running at 500MHz with 16 programmable pixel pipelines, six vertex engines and the ability to effortlessly render real time shadows as well as doing massive scientific computing, you would have smacked us silly with a trout heavier than an SGI workstation. Well, it’s 2005 and that’s exactly what GPUs can do. These breed of processors grow so fast that they’re said to defy Moore’s Law.

Of course, if a chip is made from silicon, it can’t beat Moore’s Law. GPU performance has grown faster than CPU’s but their transistor count at the same price point remains the same, in line with Moore’s prediction.

GPU performance increases have come mainly from higher clock speeds and additional pipelines. Unlike the CPU which has exhausted its tricks, future GPUs will continue to use these two techniques to drive performance.

Today’s top of the line GPUs contain six vertex processors to manipulate geometry and 16 pixel processors to work on pixels. Both types of processors can do similar work and indeed have very similar instruction sets. With the next generation of Windows, Longhorn and the introduction of Windows Foundation Graphics 2.0, pixel and vertex shaders may very well be unified.

While a common piece of hardware with the same instruction set is nice and elegant, it’s not clear that it’s the best solution. NVIDIA is not convinced. Their Chief Scientist David Kirk called such shader hardware ‘Jack of all trades, master of none.’ When we spoke to John Montrym, NVIDIA’s Chief Architect, he also gave a lukewarm assessment of unified shader hardware. ATI on the other hand has made it clear that its future GPUs will employ unified shader architecture.

So who’s right? It’s a delicate balancing of issues. A specialised vertex or pixel shader will always be faster than a ‘generalised’ one. So individually, unified architecture will be at a disadvantage. On the whole, however, a GPU using generalised shaders may prove to be more efficient than one without.

Because pixel processing can only occur after geometry processing, if one of the two stages takes too long, the other sits idle. For example, a vertex shader intensive game with simple pixels operations will overload the six vertex engines while leaving much of the 16 pixel pipelines sitting still. Conversely, a game that’s low in geometry but applies a dozen effects to everything will choke the pixel pipelines while leaving the vertex hardware unused.

Unified shader hardware won’t have such a problem. If a game is very well balanced between vertex and pixel shading, a unified shader GPU with 32 general shader units will split them evenly for vertex and pixel shading. If a game is geometry dominant, more units will be allocated to do just that while the remaining will work on pixels. As a game’s content changes from frame to frame, such a GPU will be able to intelligently allocate its shader resources to best draw the picture.

While such a scheme sounds almost too good to be true, the alternative is by no means doomed. NVIDIA has a track record of making very efficient hardware. Current architectures have fragment buffers between the vertex and pixel pipelines. This alleviates much of the work balancing problem by providing a constant pool of pixels for the pixel shaders to work on. NVIDIA will most likely provide the same instruction set for both vertex and pixel shaders in future GPUs but still use different hardware for both. That being said, in the very long run, NVIDIA may eventually move to a unified architecture.

Differing implementations aside, both NVIDIA and ATI’s GPUs released alongside Longhorn will have to support Shader Model 4.0, requiring a unified instruction set across the shaders. That means the type of operations and the limits on what can be done with shaders will be the same for pixel and vertex shaders. From then on, the programmer won’t have to think about pixel or vertex instructions, just shader instructions.

Shader Model 4.0 brings in more programmability for the GPU as well as other benefits. It’s expected that the GPUs will move to support virtual memory which will give the programmer an almost unlimited space to work with, for both textures and geometry. Better higher ordered surface support and tessellation is also a big part of the next DirectX. Together with the new shader model, this will finally give programmers what they’ve been craving for, from dynamic geometry generation to curved surfaces.

In the coming years, PCs with multiple graphics boards will become even more prevalent. NVIDIA has a very clear vision on these types of systems. Their CEO dubbed upcoming PCs armed with multi-core CPU and multiple graphics cards ‘gaming supercomputers’. For them it’s the creation of an entire new market through nForce chipsets and SLI. For us it’s the opportunity to build machines twice as powerful -- or more – than ever before. And seeing now that vendors like ASUS and Gigabyte are making dual-chip boards, don’t be surprised if two dual-chip boards sit next to each other to allow quad-GPU gaming goodness.

What:  ATI R5xx series
When: 2005 - 2006

With unified pixel and vertex shader hardware, it’s a major milestone for GPU architecture. The number of shader units is anyone’s guess. Expect massive shader throughput and the bandwidth to match. A specialised version, possibly with on-die memory will power the Xbox Next.

What: NV5x series
When: 2005 - 2006

NVIDIA's Michael Hara said that NVIDIA's next generation GPU won't be architecturally that different from their current GeForce 6 which was designed to last two to three years. But we do know that a variationof its next generation GPU will power Sony's next Playstation, and Sony is known for demanding insanely powerful processors.