Super GPUs

By Doug Barney

Clustering GPUs is cool, but what if you could massively increase the power of the graphics processor itself? That’s what NVIDIA (Santa Clara, CA) has in mind with its new Tesla line.

Named after the Serbian creator of AC/DC (not the rock band) and a pioneer in wireless communications, the Tesla family of graphics processors promises to transform an ordinary desktop into a supercomputer, the company says. Each Tesla includes some 128 parallel processors.

“By dramatically reducing computation times,” said NVIDIA CEO Jen-Hsun in a prepared statement, “in some cases from weeks to hours, NVIDIA Tesla represents the single most significant disruption the high-performance computing industry has seen since Cray 1’s introduction of vector processing.”
 
NVIDIA has three ways to serve up Tesla processing. On the high end is the NVIDIA Tesla GPU Computing Server. This box can be fitted with as many as 8 Tesla processors. NVIDIA also has the Tesla Deskside Supercomputer. This PC add-on has two Tesla processors and can give the PC eight teraflops of graphics power. On the lowest end is an add-in board so PCs can be equipped with multiple Teslas.

Custom programs can be written in C using the NVIDIA CUDA development platform, which supports parallel processing and runs on Linux and XP.