Not content to let AMD steal attention with the surprisingly good performance of its latest graphics cards, Nvidia has shared details of its upcoming graphics processor design, codenamed Fermi. With no new consumer graphics products in sight for the next few months, Nvidia's focus has been on promoting its proprietary CUDA system for accelerating CPU tasks on the GPU.

Continuing the progression of graphics chips into the general CPU space, Fermi chips will accelerate general-purpose computing applications even further. The chip would include up to 512 "CUDA cores", the new name for programmable shader units. Current GTX-series GPUs have 240, and we can safely assume Fermi's will have improved capabilities as well. Memory bandwidth will be 384 bits, and fast GDDR
5 memory will keep feeding the core with information to process.

Other headlining features are the design's ability to process ECC (error correction), and double precision arithmetic for data-intensive applications such as simulations. Parallel caching helps reduce latency by intelligently arranging data in the L1 and L2 caches. Fermi designs are already planned for massive supercomputing deployments.