CNS-1 Architecture Specification

Preface In the past two decades, the fields of VLSI systems design and massively parallel computation have grown into mature disciplines. Both fields began as research topics in industrial and academic laboratories; today, they form the core technologies for several large corporations. This report proposes a massively parallel computer, the Connectionist Network Supercomputer (CNS-1), which leverages off these fields. By targeting the computer to connectionist networks and related applications, we can focus on custom chip design and efficient software to achieve performance goals which challenge the best that industry has to offer. Why are we building such a project in an academic environment? Three different issues have motivated us to design and construct the CNS-1. The first issue involves the difficulty in applying neural network techniques to very large problems. Our research in the area of speech recognition has shown that current neural algorithms are helpful in solving problems with up to one million parameters. Performing this work cost-effectively required the construction of a smaller custom neurocomputer. Will the same algorithms work for one billion parameters? To answer this question requires a system that can train very large networks one thousand times faster than our current machines. Commercial machines with this capability do not exist. The second issue relates to the field of computer architecture. A strong distinction, apparent in the publications of the field, exists between projects that are actually realized and those which are merely paper designs. As faculty, staff and students at a leading school of computer architecture, we believe it is important to be involved in the design and construction of applied software and hardware systems. The third issue is the application of VLSI chip technology to system design. Dealing with the complexity of competitive VLSI design is a daunting challenge for an academic research group. However, several factors have led us to include custom VLSI design as a key element in the CNS-1 project. Technically, only custom VLSI lets us exploit the simplicity and parallelism of connectionist computation. From an academic standpoint, we believe university VLSI programs must undertake design projects that meet and challenge the industrial state of the art, and that these projects must be pushed to fabrication, test, and use in prototype systems. This approach teaches students relevant perspectives on VLSI design, and results in significant contributions to the engineering literature. The goal of the CNS-1 project is to produce a system …