Nvidia and IBM co-developed NV Link, a chip-to-chip interface aimed at linking GPU and CPU modules in a chassis. It will deliver rates up to 80 Gbytes/second, far above the 16 GBytes/second rates of PCI Express 3.0.
Nvidia will first use it in 2016 systems with its Pascal GPU. IBM did not say how or when it will use it.
Neither company provided any technical details or even information on when they will become available. An IBM representative said its implementation of NVLink should ultimately be available through its Open Power Consortium.
Jen-Hsun Huang, Nvidia's CEO, said a second generation of the interface is already in the works, doubling to eight the number of channels NVLink will support. The cache coherent link will let CPUs and GPUs share memory.
NVLink can connect GPUs to each other or to CPUs.
In some ways, NVLink is an alternative to processor buses such as the Quick Path Interconnect Intel uses to link to co-processors. NVLink can also be seen as an upgrade of the SLI or Crossfire technologies NVidia and archrival AMD have used to link multiple GPUs.
Next page: Nvidia rolls Titan Z