Nvidia's Fermi Core Revealed Page: 1
Nvidia have today announced their Fermi core architecture, which will form the basis of their upcoming GeForce GT300 series of graphics cards, amongst others. This is the first major design change since the launch of the G80 core in 2006. Fermi, presumably named after the Italian physicist who is credited with the invention of nuclear reactor. It now makes sense as early rumours stated the GT300 board was also called "reactor".
Fermi packs in 3 billion transistors, 512 CUDA cores (renamed from shader cores) and a 384-bit memory interface into a 40nm GPU. Also in there is 1MB of L1 cache and 768KB of L2 unified cache memory, to speed up access times. We can expect up to 1.5GB of memory on the GeForce models, with up to 6GB of memory on the high end Tesla options. The Fermi core supports Half Speed IEEE 754 Double Precision as well as C (CUDA), C++, DirectCompute, Fortran, OpenCL, OpenGL 3.1 and OpenGL 3.2. This shows how Nvidia's latest GPU offerings are aimed towards the computing and GPGPU side of things rather than solely gaming.
NVIDIA's "Fermi" architecture will be released in three tiers, GeForce for high-end consumers, Quadro for commersial use and Tesla for scientific purposes. There is still no word on pricing or availablility, but I'd expect the new Fermi cards to break the bank compared to ATI's offerings.