Mr. Steve Teig

President and CEO of Tabula

Abstract

The "von Neumann architecture" for computers, invented mostly by Turing but popularized by the more famous von Neumann, has completely dominated computing for more than 65 years. It is a masterpiece of simplicity: readily implemented in hardware, easily understood by software developers, and amenable to compilation from a wide variety of programming languages. Unfortunately, it achieves its simplicity from the fundamental, non-physical assumption that reading from a memory location takes negligible, constant time independent of the size of the memory. Decades of innovation in computer architecture and compiler design for uniprocessors has masked some of the von Neumann computer's intrinsic latency. The power requirements for this disguise have become prohibitive, though, which has ended the long, exponential rise in uniprocessor clock frequency. Multi-core processors, the semiconductor industry's response, have the virtue that they can clearly be built, but no one knows how to program them! Further, they make the same negligible-latency assumptions as uniprocessors, but disguising that latency is now quadratically more difficult.

This talk will show that highly useful yet non-physical oversimplifications such as the von Neumann architecture have numerous historical precedents from which we can learn. These examples suggest that a more physically aware, non-von Neumann machine could offer higher-performance and far more power-efficient computation. Next, we offer some thoughts on what such a machine might look like - hint: it is not an array of microprocessors! - and how one might program it. It is only by simultaneously approaching architecture, hardware, and software, seeing them as aspects of a cohesive whole as von Neumann and Turing both did, that we maximize our chances of going beyond von Neumann computing.