SemiAccurate is reporting that NVidia will be releasing Fermi with 448 cores instead of 512 cores because it is too hot to handle. In the past, SemiAccurate has lived up to its name with the false rumor that ORNL canceled its Fermi Supercomputer. This time, however, they link to a TESLA C2050 and TESLA C2070 Computing Processor Board PDF. The Tesla C2050 and C2070 are both based on the Fermi chip. The document shows that the Tesla C2050 will likely come with 448 cores, running at 1.25 GHz, and have 3 GB of RAM. The Tesla C2070 will likely come with 448 cores, running at 1.4 GHz, and have 6 GB of RAM.
Back on September 30, 2009, Nvidia released a white paper describing the Fermi architecture. Fermi was to have 16 streaming multiprocessors. Each streaming multiprocessor had 32 cores. In total, Fermi would then have 512 cores.
However, this PDF document is more recent, with a release date of November 16, 2009. It states that Fermi will have 14 streaming multiprocessors. In total, Fermi would then have 448 cores.
It is also important to note that NVidia does include the disclaimer in the Fermi Compute Architecture Whitepaper that:
The first Fermi based GPU … features up to 512 CUDA cores.
It was originally thought that the top-of-the-line Fermi card would have 512 cores, while a variant might have fewer cores. For example, the 280 GTX has 240 cores, while the 260 has 216 cores, which itself is up from an initial 192 cores. It is still possible that NVidia might release a 512 core variant, perhaps once the manufacturing process improves.
It is also possible that the gaming card will have 512 cores, while the Tesla will have fewer cores. It has often been the case that NVidia releases a gaming card that has higher clock speeds and memory capacity than the equivalent Quadro or Tesla. However, one would think that since a Tesla costs more than the gaming card equivalent, that NVidia would rather earn a higher profit margin on the Tesla rather than put it in the gaming card. Either way, the debate will rage on until NVidia actually starts selling the product, which should be in March 2010.
It would seem the logical step, nvidia really need to get something out into the market, the longer they delay getting their first dx11 cards out, the longer ati will have had refining theirs. Unless fermi turns out to be one hell of a gpu (and i’m talking +50% performance on the hd5870) they are going to really struggle this generation.
They may (and should) take the performance crown, but i would think ati could probably halve the price of the hd58xx series without breaking a sweat since the chip is so small. Whereas fermi is going to be a lumbering behemoth and will cost a helluva lot (I predict £500 launch price) and will have small margins even at that price.