At his keynote address at GTC 2012, NVIDIA CEO Jen-Hsun Huang announced two new Tesla GPUs based on the Kepler architecture. They are Tesla K10 and K20 respectively. The former boasts of performance improvement of up to 3x over Fermi-based Tesla GPUs for single-point floating point operations while the more powerful Tesla K20 boasts of similar gains but with regards to more computational-intensive double-point floating operations over the previous generation Tesla members.
Hence, to their different computational capabilities, Tesla K10 is targeted at industries that need to process data from radio signals or seismic data from oil and gas exploration. Industries with supercomputing requirements will look to the Tesla K20 that will be made available in the fourth quarter of this year. With that said, we have put to rest our earlier speculation about the possible launch of a "full-fledged" version of the Kepler GPU.
During the Q & A session after his keynote address, CEO Huang did confirm that the Tesla K20 features a 384-bit wide memory bus but he didn't mention the amount of onboard video memory that will ship with the K20. This is because of their dependency, like any other graphics card manufacturer, on the DRAM industry to what lay beyond the current GDDR5. With regards to the pricing of these cards, CEO Huang didn't commit to an answer; however, we believe that the high-performance computing and other related industries would rate NVIDIA's commitment to making GPU-accelerated computing easier and more convenient as priceless.