Improbable Insights

Page :« 1 2 3 4»

Then There’s CUDA

Finally, let’s consider CUDA.

CUDA, or Compute Unified Device Architecture, is Nvidia’s enhanced version of what began as GP-GPU (general purpose GPU) computing. CUDA allows developers to take advantage of the inherently massively parallel floating point engines built into modern GPUs. The GT200 line was specifically enhanced to make life easier for people who want to write non-graphics applications that run on GPUs.

AMD has a similar initiative with the Radeon line called Stream Computing. From what we’ve seen, though, CUDA seems to be a more full-featured platform and early performance numbers seem to favor Nvidia. The problem here is that the standards world is catching up with Nvidia. OpenCL is now a reality, and DirectX 11 will soon be upon us and with it, DirectX 11 compute shaders.

And on the horizon looms Larrabee. Larrabee is Intel’s first stab at a high end GPU, built around a bunch of slimmed-down, Atom-like x86 cores with significan vector instruction extensions. Larrabee will also feature other features in common with general purpose CPUs, like relatively large caches and cache coherency. Almost all of Larrabee’s graphics functionality will be programmable, and focused on a tile-based rendering approach.

To really damage Nvidia (and, by extension, AMD) Larrabee only needs to be a modest success. If it achieves graphics performance only equal to today’s GTX 285 (plus DirectX 11 support), but demonstrates massive improvements in compute functionality, then developers will probably favor Larrabee over the competition.

Interestingly, if Larrabee does inflict serious damage, Nvidia may have sown the seeds of that result themselves. I’ve heard from several people in the know that Intel would never have pursued Larrabee had Nvidia not banged the drum for CUDA so hard. As Isoroku Yamamoto observed during a completely different conflict, “We have awoken a sleeping giant.”