Nvidia is diversifying into HPC with a new family of high-margin graphics chips called Tesla, a counter to ATI's stream processor card that's supposed to turn existing workstations into "personal supercomputers."

The company claims it represents the "single most significant disruption the high-performance computing industry has seen since Cray 1's introduction of vector processing."

My. My.

It's a co-processor with 128 parallel processors that will work with high-performance CPUs. Nvidia advises one CPU core per GPU core.

The chip is supposed to be able to deliver 518 gigaflops of parallel computing and it's generally conceded that GPUs are more efficient at parallel processing than general-purpose processors.

Nvidia is packaging the widgetry, which is a redirection of its existing 8-series GPU, as a PCI Express x16 add-in card, a two-Tesla desk-side box (figure 1 teraflops), and a cluster-aimed 1U server that will ultimately house eight of the little beasts, or 1,000-odd parallel processors, four to start (2 teraflops). This last is supposed to be a first of a kind for the data center.

Tesla uses Nvidia's CUDA C development environment that's supposed to simplify parallel computing and the hassles of massive multithreading on a GPU because it's standard C and can process thousands of threads simultaneously.

It includes a compiler, debugger/profiler, dedicated driver and standard libraries and is supported on Linux and 32-bit Windows.

Currently Tesla can only do single precision floating point. Double is due at the end of the year.

The D870 card runs $1,499, the Quadro Plex-reminiscent C670 box $7,500, and the S870 server $12,000. No rush though, the card and the box won't be available until August and the server until November or December.

The market Nvidia is chasing is supposed to be worth over $6 billion by 2010. By then Intel will have joined.