Dr. Eugen Schenfeld, T.J. Watson Research Center, Yorktown Heights, NY USA

The technology and architecture for large systems' networks has come a long way with a plethora of network topologies, technologies and protocols being studied and proposed in the past. As systems become more powerful and distributed, it is clear that the network is a key component. For example, many studies have shown that the trend of the bytes/flops measure, indicating the ratio between network and processing power in a system, is getting worse with time, where processors get more powerful yet the bandwidths and latencies of the network connecting them get worse.

In this talk I will review a novel trend emerging as a possible solution to build the next generation of powerful networks. Networks that could offer a wire-once, 0-latency (up to speed of light latency limit), attractive cost, yet able to handle and be used through several generations of processing systems and link protocols. The origin of such networks goes back to the late 80s. Recently, this new trend has been more widely adopted, with multiple examples of work by various academia and industrial teams as recent as in the last 2 years or so. I will provide example of use and enabling technology from both HPC area as well as for more commercial datacenters.