The Fermi architecture is a dream come true for GPGPU applications and NVIDIA has worked hard on the server and workstation performance. Even if the Fermi architecture that was presented back in 2009 is the same, the exact specifications released to partners have been altered quite a bit. Numbers from three different specifications have been published and they don’t paint a pretty picture of Fermi Tesla.
Since the first specifications was sent out in May 2009 it was first updated in November 2009 and again in April 2010 when the cards finally appeared.
Tesla C2050/C2070 is starting to find its way into servers, e.g. from Supermicro
A lot has happened during the last year, and the fact is that NVIDIA’s Tesla C2050/C2070 graphics card only delivers 68% of the first promised performance of the Fermi architecture. The reason is the reduced number of CUDA processor, from 512 to 448, and clock frequencies of both CUDA processors and graphics memory was lowered several times.
To make things worse power consumption has increased by 41% from the original 175W to 247W.
What this means for Fermi Tesla’s performance relative to competing solutions is still a bit foggy, but there is no doubt NVIDIA has had to settle after running into too may obstacles.