A French mathematician names Nvidia’s next graphics architecture. At the GPU Technology Conference it is announced that the successor to today’s Maxwell will take over in 2016 and unlike in last year’s presentation is not called Volta but is now known as (Blaise) Pascal.
The new architecture will deliver several innovations. First and foremost, stacked memory circuits are added with promises of many times higher memory bandwidth than conventional techniques. Laying circuits on top of each other should also reduce energy consumption and make it possible to expand with 2.5 times more memory.
Another new addition is the data bus Nvlink with 5-12 times higher bandwidth than today’s PCI Express 3.0. The purpose is to speed up the communication between GPU and CPU, something that is especially useful in supercomputers and high performance computing (HPC), but can also provide a performance boost in traditional PCs and future incarnations of SLI.
At the presentation, Nvidia’s CEO Jen-Hsun Huang showed a modular version of Pascal, which is apparently intended for mainframes. The module is only a third as large as a traditional PCI Express-based plug-in card and appears to contain a graphics circuit surrounded by memory capsules and voltage regulation.
According to Nvidia’s planning, the first graphics processors with the Pascal architecture will appear sometime in 2016. This also means that the newly launched architecture Maxwell will also be relevant in 2015.