updated 02:20 pm EST, Mon November 16, 2009
NVIDIA 20 series Teslas appear
NVIDIA today provided details of the first official hardware to use its upcoming Fermi architecture. The Tesla 20 series is even more optimized for general-purpose computing standards like OpenCL or NVIDIA's own CUDA and handles complex math that previously hasn't been as practical, such as ISO standard double-precision math and C++ code processing. Unlike past models, though, the card model also has a video output and works as a video card rather than just as a companion device.
The C2050 and C2070 fit into PCI Express 2.0 slots and respectively carry 3GB or 6GB of GDDR5 memory to hold most of the data on the card instead of resorting to system RAM; NVIDIA estimates as much as 630 gigaflops of peak performance. The S2050 and S2070 are 1U rackmount units that link up four C2050 or C2070 chipsets in a single system and push up to 2.5 teraflops with either 12GB or 24GB of GDDR5.
NVIDIA doesn't plan to ship the new Tesla range until spring of next year but already prices the C20-series cards at $2,499 and $3,999 each, while the S20 units cost $12,995 and and $18,995 respectively. Operating system support hasn't been detailed at this early stage.