It is another confirmation that the accelerators NVIDIA on GPU computing Volta will appear in the next year. Previously, the company alternated graphics architecture in increments of two years. Now, after Pascal architecture, presented this spring, next spring the company will bring to the people a new architecture called Volta. It will be based on 16-nm process technology and the first time will make a couple of processors IBM Power9. In the consumer segment GPU-based Volta is unlikely to fall until 2018.
According to sources in the GPU Volta enters at least four ports NVLink second generation. In the first generation NVLink each port provides two-way exchange of data at speeds up to 40 Gb / s. As part of the GPU P100 four such ports, which makes the rate of exchange at the level of 160 GB / s. NVLink 2.0 interface to obtain an increase in the exchange rate of 25%, or 50 GB / s per port. In the case of four-port configuration, as is done in the GPU P100, the rate of exchange will increase to 200 GB / s. This type of interface is a cellular network, which communicate between the GPU and a CPU (processor will incur if built NVLink interface).
NVIDIA's highly appreciated NVLink interface and NVLink 2.0. According to the developers - is an uncompromising alternative to the PCI Express bus. The rate of exchange for the one-slot PCI Express 3.0 x16 reaches 32 GB / s, and even in the case of PCI Express 4.0, which will appear next year in the x16 configuration exchange rate does not exceed 64 GB / s. Block NVLink structure in this respect is more flexible and forward-looking.
IBM, in 2017 will begin to unfold in the United States Summit system at the Oak Ridge National Laboratory and the Sierra system at Lawrence Livermore National Laboratory. Lawrence. Summit Peak performance will be 150-300 petaflops, and Sierra Systems - 100 petaflops. All of them will be borne by IBM Power9 processors paired with the accelerators NVIDIA Volta.
Related Products :
|