Nvidia presents new GPU interface and architecture

Nvidia plans to build a new, faster interface into its GPUs so that GPUs and CPUs can share data five times faster than today. The company will also launch a new series of GPUs in 2016, under the name Pascal.

According to Nvidia, the current PCI Express 3.0 interface that is used as an interface to insert a GPU into a PC is an important bottleneck. The NVLink interface should solve that, Nvidia announced during a press conference. NVLink is five to twelve times faster than PCI Express 3.0. The caches between the CPU and the GPU also remain synchronized.

According to Nvidia, being able to move data back and forth more quickly is especially important for supercomputers. According to the company, the new interface will enable supercomputers with 1000 petaflops, which is 50 to 100 times faster than current supercomputers.

The new interface will still be applied with Nvidia GPUs that use the Maxwell architecture, but when exactly is not yet known. The first GPUs with that architecture were announced last month.

In addition, Nvidia will release GPUs with a new architecture in 2016, under the name Pascal. That architecture must be equipped with memory that is placed on top of the gpu, so that a gpu can work with thousands of bits at the same time, where that is currently a maximum of hundreds of bits. That principle, called 3d memory by Nvidia, also ensures smaller high-end GPUs: during the keynote, the CEO of Nvidia showed a Pascal card the size of a debit card.