Kirk, delivering a chalk talk on "The Future of 3D Graphics' in San Francisco Friday, touted the advantages of GPU-based parallel computing for powering applications related to oil and gas exploration, computational finance and other computational modeling projects, as well as for faster, more powerful hybrid rendering within the graphics discipline itself. Because GPU computing is already a "data parallel process," the work of breaking apart computing problems into smaller sets of instructions to be carried out concurrently is more easily done on GPUs than on multi-core CPUs, Kirk said.
Describing a kind of Moore's Law on steroids, he promised 100x performance gains in real-world applications, just as soon as people take advantage of Nvidia's General-Purpose computing on GPUs (GPGPU) initiatives that have resulted in some 50 million Nvidia GPUs already shipped that are capable of running the CUDA programming language for parallel computing.
"This is truly the democratization of supercomputing. We ship a million parallel units a week," Kirk said.
CUDA or Compute Unified Device Architecture, is a C programming language developed by Santa Clara, Calif.-based Nvidia that allows GPGPU programmers to code algorithms for execution on graphics processors. Currently, it's possible to run CUDA on Nvidia's GeForce desktop chipsets, as well as its Quadro workstation and Tesla high-performance compute products, and according to Kirk the graphics chipmaker recently released an SDK for the Macintosh operating system.
NVIDIA: CUDA to deliver 100x performance gains
Posted on Monday, February 25 2008 @ 0:15 CET by Thomas De Maesschalck