NVIDIA have been spilling the beans on the Fermi GF100, the company's 40nm GPU with a full 512 CUDA cores, which they reckon will deliver not only exceptional gaming performance but film-like geometric graphics. To do so, they've thrown out much of the traditional graphics pipeline structure and introduced far more parallel geometry processing; in fact the GF100's GigaThread engine can flip between contexts - DirectX 11 for rendering, CUDA for selective ray tracing, Direct Compute for post-processing and PhysX for simulations - in around 20 microseconds.
The GF100 has four Graphics Processing Clusters (GPCs), sixteen Streaming Multiprocessors (SMs) and six GDDR5 memory controllers (for a 384-bit total); each SM has 32 CUDA cores for a total of 512. Meanwhile each GPC has a raster engine and up to four SMs, and in fact each GPC could be considered pretty much a self-contained GPU in its own right.
There's plenty more technical detail in NVIDIA's whitepaper [pdf link], but in the end it's performance that most would-be buyers are interested in. According to PC Perspective, the GF100 certainly has the potential to beat out single-GPU cards from ATI, such as the HD 5870, perhaps by a 15- to 20-percent margin, though it might fall short of a dual-GPU card. They also offer up the following two videos; first of the GF100 rendering Far Cry 2 and second of the GF100's ray-tracing performance.