Nvidia's CUDA

Nvidia definitely weren’t the first on the scene. A few years ago, ATI announced a hardware interface now known as Close To Metal (CTM). While hardware support is possibly more advanced, the CTM framework appears to be very low level and challenging to work with. This shortfall is where CUDA shines by providing numerous tiers to access the GPU, from simple C extensions down to assembly language.

At the beginning of the year, Nvidia announced a takeover bid for Ageia --most famously known for PhysX or ‘that physics card that no one bought’. Ageia’s pride and joy wasn’t the card, it was the software of the same name designed to provide a standardised physics engine. The card only accelerated this software with its parallel processor, which was similar to a GPU.

Nvidia’s recent announcement of PhysX for Geforce GPUs means the PhysX physics engine will take advantage of your GPU in the same way CUDA does. This means suddenly physics acceleration is back on the table, and doesn’t need you to part with any extra moolah for the privilege.

Nvidia aren’t just marketing CUDA for the gamer demographic though. It’s also launched its Tesla card, which contains essentially the same hardware as a GeForce 8800 GTX, but isn’t a video card at all. It’s a budget super computer, not a graphics display. And it’s doing science. And is still alive.

So where’s triple – or even quad -- SLI during all this? It may not be terribly beneficial to gamers, but SLI but does have the upside of moving more video cards out the door of Nvidia HQ. Unlike games however, when you’re writing applications with CUDA that utilise umpteen-thousand threads, a doubling, tripling, even a theoretical n-fold increase in the capacity of the processor generally scales performance accordingly.

Intel has downplayed all of this, stating that GPGPU languages such as CUDA will only be “interesting footnotes in the history of computing annals”. This isn’t really surprising, considering Intel has its own Cell processors to push, and Nvidia could one day directly threaten Intel’s entire CPU business. Intel also owns PhysX’s primary competitor – the Havok engine. So take Intel’s derision with a grain of salt.

CUDA will only be a footnote if there is no adoption by mainstream programmers. Given the amount of resources Nvidia is pouring into educating programmers, including running courses in major US universities and colleges that give credit toward degrees, it’s not going to die without a fight.

Will mainstream games actually run with this stuff? We think so. We’ll likely see a number of tech-demos before real games fully realise the potential of GPGPU, but it will happen.