|
Memory bandwidth is becoming even more important feature in modern GPUs. It is the reason why today new 3D Memory and High Bandwidth Memory (HBM) technologies are paid closer attention to. Back-end systems experience a similar problem of data transfer to and from GPU and other components of the system. At GTC 2014, NVidia introduced NVLink new technology, which is capable of providing high connection speed in the nearest future.
NVLink’s PCI Express bus bandwidth is up to 16 GB/s; the speed between memory and processing unit amounts to 60 GB/s; modern GPUs work with memory by 512-bit bus at the rate of about 300 GB/s. In spite of the fact that some technologies that can enhance speed of interaction with memory has been already announced, most producers still have to resign to certain technical limitations; therefore, they are enforced to look for the solutions, which will make the components interact with each other in the systems of parallel processes of a high level.
GTC 2014 saw a back-end NVLink version of Pascal architecture. NVidia NVLink interface is based on point-to-point direct connections, which can be combined with each other differently. For example, it can be a simple GPU-CPU connection or a complex one as GPU-CPU and GPU-GPU. Each connection consists of eight lines. Initially, Pascal will use four NVLink connections. According to the report, performed by NVidia, number of the connections varies according to a target market. Back-end subdivision at IBM is interested in this interface. The companies have informed the public about their collaboration.
Recently, IBM represented Power S824L system, which uses IBM Power 8 processors and Tesla graphic cards. As potential ways of application, IBM regards systems of databases and GROMACS, NAMD, SOAP3, FFTW and Quantum Espresso models computation. However, there are still some constraints that are planned to be eliminated by NVLink interface since 2016. It will not only combine several GPUs together, but will increase data transfer rate among them, thus the server’s efficiency on the whole.
However, NVidia and IBM do not specify, how NVLink interface will look like. More details may be revealed at GTC 2015; also, then it will become clear, whether end customers are interested in it. |
