nVidia has now managed to put together what they seem to think is a reference design for an "affordable" (roughly $10,000 US) deskside supercomputer. Based on the second generation Telsa card, the new suberbox would contain four C1060 Tesla cards and need at least a Quad Core CPU and 16GB of RAM (nVidia recommendation).
TGdaily has more information here.
If you are into numbers, such a PC would have 960 graphics cores and a combined performance rating of about 3.6 to 4 TFlops in single-precision and about 400 GFlops in double-precision applications. The C1060 GPUs (T10P processor) are clocked at 1.33 GHz and rated at a performance of about 900 GFlops in a single-unit configuration.
Depending on the application performed, Nvidia claims that a Tesla PC will be about 250 times faster than a regular desktop PC. However, we need to be fair and mention that this number also depends on the system the PC is compared to. For example, a PC equipped with an Nvidia SLI system or two of ATI's Radeon HD 4870 X2 cards can compete with a Tesla PC in the Teraflops department. In fact, ATI says its tow 4870 X2 cards provide about 2.4 TFlops in single-precision and will actually outperform the four Tesla cards in double-precision.