NODE INTERCONNECT ARCHITECTURE TO IMPLEMENT HIGH-PERFORMANCE SUPERCOMPUTER

摘要

Node interconnect architectures to implement a high performance supercomputer are provided. For example, a node interconnect architecture for connecting a multitude of nodes (or processors) of a supercomputer is implemented using an all-to-all electrical and optical connection network which provides two independent communication paths between any two processors of the supercomputer, wherein a communication path includes at most two electrical links and one optical link.

主权项

1. A computing system, comprising:
a plurality N of computing groups which are optically connected to each other to form a computing system, wherein each computing group comprises: a local group of N multi-processor modules; a local group of N optical redistribution boxes; a plurality (N×N) of local optical bundles which optically connect the local group of N multi-processor modules to the local group of N optical redistribution boxes; and a plurality (N×N) of global optical bundles, wherein N of the global optical bundles optically connect the local group of N optical redistribution boxes to the local group of N multi-processor modules, and wherein (N×N)—N of the global optical bundles optically connect other local groups of optical redistribution boxes in other computing groups to the local group of N multi-processor modules.