The first is the Tesla C870 "GPU Computing Processor," which looks strikingly similar to the Quadro FX 5600 that came out three months ago. Just like the Quadro, the Tesla C870 comes with 1.5GB of onboard GDDR3 memory clocked at 800MHz. Its processing units run at 1.35GHz, and Nvidia claims peak floating point power of over 500 billion floating point operations per second (gigaFLOPS). The card slips into a standard PCI Express interface, although unlike AMD's Stream Processor, it has no display output ports.
Nvidia's second Tesla solution is the Tesla D870 "Deskside Supercomputer," which is a version of Nvidia's Quadro Plex VCS Model IV multi-GPU enclosure built to house a pair of Tesla C870 cards. As with the Quadro Plex, two Tesla D870 enclosures can be paired up to fit inside a 3U server rack space.
Last, but not least, is Nvidia's Tesla S870 "GPU Computing Server." This is in essence a D870 on crack that was designed for server racks from the start. It fits inside a 19" 1U rack-mount space and houses no fewer than four Tesla C870 cards, making for maximum combined floating point computing power of over two teraFLOPS. Nvidia says power consumption for the S870 is 550W typically and can go as high as 800W.
Of course, that's just the hardware. As one might expect, Nvidia allows developers to tap into Tesla hardware with its CUDA application programming interface. Our counterparts over at Beyond3D have a more detailed outline of the Tesla lineup and how CUDA comes into play over in this article.