Nvidia prepares to spoil AMD's GPGPU party

THE EYES of AMD/ATI fans this week were trained on the hot Tunisian capital, for a drip of at least some exciting news.

Yes, there's Tatooine there, but AMD is no Luke Skywalker any more, and Intel's Darth Vader has repented with more friendliness in from the positive side of the Force. So, the scary NDAs (almost) keep most of the obliging sheepish "press" from reporting anything useful from Tatooine, while the rest of the PC market roaming giants do their job.

One of the biggest among the giants, Nvidia, has been busy protecting its high-margin OpenGL turf, and seeking to prove it can compete in the emerging GPGPU market. That's where some venture capital money is coming in - RapidMind, another GPGPU software layer vendor, just got another US$10 million of VC money this week.

So, Nvidia updated its Quadro Plex visual computing system (VCS) with the additional GPU computing resources of two of its brand new Quadro FX 5600 GPUs - read OpenGL spin of 8800GTX with twice the memory, 1.5 GB per card, but a bit slower.

The expensive teraflop-class external box, connected to the PC via PCI-E x16 external link, is aimed at workstations and cluster in styling, design, oil and gas, and scientific apps. To make use of the GPU computing, you have to master Nvidia's CUDA programming - might still be better than assembly-coding the Radeon X1900, though.

On the OpenGL side, Nvidia still has some advantage over AMD, as it focused all these years on perfecting the application support and shelling out the dosh to all those greedy application vendors for "certification expenses". On top of that, its OpenGL software optimisation is also quite a bit ahead, unless AMD takes a serious note of this market's importance and fixes the problem - which it is, slowly, doing now in all those monthly Catalyst updates.

Finally, Quadro Plex claims some nasty 3-D performance numbers, even for an external box - up to 80 billion pixels per second and seven billion vertices per second, and able to handle up to 148 megapixels display walls or 3-D caves on 16 synchronised digital-output channels or eight HD SDI channels. This is achieved using Frame Synchronization where display channels from multiple workstations are combined into one large virtual display - imagine immersing yourself in a photorealistic virtual 3-D room of a kind.

And, of course, you can have many of these Quadro Plex boxen in your visualisation cluster for "immense scalability", according to Nvidia. It could be an immense scalability in its profits, as each Quadro Plex VCS Model IV is priced at $17,495 per box, and more so in good old Blighty, no doubt. µ