While ATI and its newfound parent AMD continue discussing the potential benefits of actually pairing their technologies into one cohesive unit, now that their companies have been paired together in a similar fashion, their principal rival in the graphics arena decided it isn't waiting to make a similar play with Intel.

NVidia today may have launched the stand-alone GPU-centric computer business all by its lonesome, with today's announcement of a kind of computer system specifically designed to mesh graphics processors together to perform rich math functions.

The goal of nVidia's new Tesla computer, if it can be believed, is nothing short of staggering: over 2,000 gigaflops - billions of floating-point operations per second - in a system that meshes together four GPUs in parallel, each of which contains not two, not four, but 128 pipeline processors. If this is true, nVidia has surged all the way into multi-teraflop territory with a device that fits into a 1U package. By comparison, a 4P dual-core Itanium-based server (eight cores total) registers about 45 gigaflops in recent LINPACK tests.

Of course, the secret is that these GPUs aren't working in place of CPUs. They can't, because their instruction sets are not compatible. Applications have to be written in C and compiled the old-fashioned way, for execution through an operating system driver that dispatches math instructions to the GPU cluster's multiple pipelines. NVidia already produces tools for compiling C applications for GPU execution, using what the company calls CUDA architecture.

It will be hard to make initial comparisons at this level of CUDA applications against traditional CPU-driven programs, because if you think about it, a "floating-point operation" is a concept that is based essentially on CPU architectures. In the GPU realm, where matrix calculations can drive thousands of parallel computations just as easily as a single one - and with memory bandwidth hoisted to an incredible 76.8 GB/sec - a "flop" may be something entirely different than we generally accept it to be.

The first Tesla units are scheduled to go on sale this August. Don't expect it to be sold to the back-to-school crowd or to the gaming elite; this is a device that universities and research institutions will want to study. The problem they face now is, do they have to buy one in order to study it to see whether it's worth purchasing?

University of Illinois Urbana-Champaign senior research programmer John Stone has jumped the gun on his academic colleagues, and has already experimented with prototypes.

In a prepared statement today, Stone said, "Many of the molecular structures we analyze are so large that they can take weeks of processing time to run the calculations required for their physical simulation. NVidia's GPU computing technology has given us a 100-fold increase in some of our programs, and this is on desktop machines where previously we would have had to run these calculations to a cluster."

The Tesla will support 32- and 64-bit Red Hat Enterprise Linux versions 3, 4, and 5; 32- and 64-bit SUSE Enterprise Linux versions 10.1, 10.2, and 10.3; and 32-bit Windows XP (curiously not 64-bit yet, and not Vista).