NVIDIA Fermi architecture more details

Check this out, this is a table that compares NVIDIA's upcoming Fermi (GT300) architecture with the GT200 and G80 architectures. For the full details you have to check out the whitepaper (PDF) NVIDIA published on its website.

GeForce, Quadro and Tesla chipsets will jump from 240 cores to a much larger 512 and should be much faster in each core courtesy of some industry-first techniques. Fermi chips will be the first GPUs to have a real cache hierarchy, with Level 1 caches to keep specific information on hand and a single, shared Level 2 cache for larger tasks; they will also have a new GigaThread engine that can transfer data in both directions at once and handle "thousands" of tasks at once.

ECC memory will also work for the first time and should help Quadros and Teslas avoid errors for large scale chores, especially in PC clusters.

The chipmaker is short on full performance claims but says Fermi is much better tuned for general purpose computing. Its double-precision floating point math should run about eight times faster than the company's previous best and should provide much better performance with current and upcoming standards, including OpenCL, Microsoft's DirectCompute and NVIDIA's own CUDA.

Below (click on thumbs) a chart, block diagram and a die shot :)