While the PCI-Express 4.0 interface will arrive next year by offering a bandwidth of 64 GB / s (but more interesting will be the 300 watt supply to GPUs), the same year will debut the new interconnection NVIDIA NVLINK 2.0 with a bandwidth of 200 GB / s, it is more than evident why NVIDIA has not replaced the NVLink by PCIe 4.0 and prefer to release a new version. NVLink 2.0 would benefit from faster interlink between the CPU and GPU (from 25 Gbps – 20 Gbps), which translates into 50 GB/s bi-directional bandwidth.
The next Tesla V100 (model name is not confirmed at the moment) could also be designed in a SXM2 format (which basically confirms that it has high bandwidth memory). Like Big Volta (aka GV100), which could be available in two variants (computer – gaming-oriented), but still unknown. NVIDIA Volta is manufactured at 16nm FinFET process node, the same used in Pascal. If the latest rumors are correct, it could be available in May 2017.
IBM’s upcoming Power9 chip will include NVLink 2.0 technology, and servers using the chip will be available next year. For now, the NVLink 2.0 I/O is designed for NVIDIA GPUs, and no other component works with the interconnect. IBM’s Power chip also offers, the PCI-Express 4.0 connector and a CAPI (coherent accelerator processor interface) interface for other chips, storage, and memory.