The competition between NVIDIA and AMD has reached new heights – the former company just launched the GeForce GTX Titan X, while the latter is describing the Radeon R9 390X flagship video card. It seems that the upper hand is in NVIDIA, though, as the US company has not only presented a new high-end product, but has also described its next Pascal architecture to a certain extent.
Pascal will appear no earlier than 2016 but even then the architecture will bring a number of enhancements. GPUs on the Pascal architecture will support up to 32 GB of HBM memory and will provide some flexibility – cards with this much memory will either employ four 8 GB HBM layers or eight 4 GB layers for a total of 32 GB of on-board video memory. The new memory organization will allow Pascal to reach three times more memory bandwidth than what is currently possible with Maxwell. Furthermore Pascal will also have an integrated NVLink interface that will allow connecting two or more GPUs via a very fast data link. Pascal-based GPUs will be able to support up to two NVLink interfaces with up to 20 GB/sec of data transfer speed. Finally NVIDIA claims that Pascal will offer unmatched performance in 16FP and 32FP calculations, which will make GPUs with this architecture very suitable for supercomputers and home PCs. Unfortunately there is no information on what 64FP performance Pascal will offer when it comes out.
All is good but the detailed description of Pascal might mean one thing only – the expected NVIDIA Volta architecture will likely never come out.
Source: NVIDIA