Yesterday master conversational genre - Director of NVIDIA, Jen-Hsun Huang - spoke about the graphics architecture sample in 2016. This architecture Pascal, in April last year suddenly came to replace the previously expected architecture Volta.
In the meantime, let us specify that the GPU-based Pascal will be supported by 32 GB of memory type HBM (High Bandwidth Memory). Since adapters based architecture Pascal will come no earlier than the first half of 2016, they will carry on a common substrate with the GPU or four 8-GB chip vosmisloynyh HBM, or eight 4-GB four-layer chip HBM. The company SK Hynix just at this time will start mass production of both types of memory chip. Chef NVIDIA emphasizes that it will be such a design module - GPU and adjacent to a common substrate chip HBM, rather than on the adapter. This will allow the interface HBM develop the maximum capacity - three times more than Maxwell architecture allows .
NVIDIA also director recalled that GPU Pascal get built-in NVLink based partially mutated interface PCI Express. According to previous reports, the GPU Pascal composition may be incorporated two blocks NVLink, and the speed of each level will be 20 GB / s.
Finally, the company clarified that the GPU Pascal succeeds in mixed calculations, illustrating the combination of the above forecast calculations with accuracy and 16FP 32FP.