HANDS ON GUIDE

5 - Can X86 Kill Nvidia and ATI?

Not content with pushing x86 into the heart of every computer, Intel wants to push it into the heart of every GPU as well. This is due to happen with a part currently codenamed Larrabee, which is scheduled to splash down around the end of 2009.

Larrabee has changed direction since it was first conceived, with Intel’s initial drips of information seeming to imply that it would be a general-purpose processor based around ‘heterogeneous’ cores. These are multiple cores, but some of these are dedicated to distinct tasks such as video decoding. However, that has now changed and Larrabee has been formally announced as being aimed at ‘visual computing’. It looks as though there will still be special-purpose cores, but the emphasis is on providing a massive array of identical x86-compatible processing units.

The target area is that of graphics. This means that Intel wants x86 to become the new standard for graphics cards, ousting Nvidia and ATI/AMD’s way of doing things.

Intel is gunning for the limitations of stream processors used in current graphics cards. Stream processors are designed from the ground up for ‘data parallel’ operations, whereby the same operation is applied to a large set of data. This is great for operations used in the traditional graphics pipeline, and means that the processing units can be simple and lean, rather than bogged down by the full complexity of the x86 architecture.

However, that advantage is also a drawback. Stream processors aren’t good for general-purpose computing. One of the key features that they lack is flexibility when it comes to divergent tasks. Nvidia’s CUDA language, for example, needs to work on blocks of threads called warps. The entire warp needs to perform identical tasks in order to provide maximum performance. As soon as a thread needs to diverge and start doing a task different from that of the others, the advantages of parallel execution evaporate. However, divergence – which includes IF and SWITCH programming statements – is the bedrock of most standard computer programs.

Intel is trying to take advantage of that relative lack of programming flexibility by forcing a complete abandonment of the traditional graphics pipeline, whereby scenes are rendered on a per-polygon basis, and push towards global illumination models such as ray tracing. Intel would claim that this is because ray tracing will give far better and more realistic results, but detractors argue that it’s because ray tracing requires lots
of divergent operations. This is perfect for an x86-based system, but not so good for a traditional graphics card.

Nvidia’s top man, Jen-Hsun Huang, launched a stinging attack on Larrabee in April, dubbing it ‘Laughabee’. Aside from being possibly the worst joke in the world, it also betrays what seems to be a genuine fear about the impact x86-based computing could have on the positions of both Nvidia and AMD/ATI. An additional problem for both Nvidia and AMD/ATI is that the tools they’re producing for flexible graphics card programming, outside the limits of DirectX, are pretty young. Nvidia’s CUDA language and AMD/ATI’s CTM are both in their infancy.

Latest Hands on guides

These are the most popular features articles over the past three months.

Mobile Broadband

Compare prices

Fastest, cheapest 3G mobile broadband dongles from 3, Vodafone, T-Mobile and Orange
from just £10/month