XPS: FULL: Emerging Nonvolatile Memory for Analog-iterative Numerical Methods

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

A new type of computer memory - crosspoint resistive memory - has emerged as a likely candidate to replace current memory technology in future computing systems. This memory allows for potential computer designs with high memory capacity and with memory incorporated directly into processing units. Novel thinking about computational methods is required to exploit the potential of these novel systems. This project will explore new, fundamental methods in the field of numerical optimization that are suited to implementation on computer systems that incorporate crosspoint resistive memory. The field of optimization is chosen as a testbed because of its importance to a wide range of scientific disciplines. Crosspoint memory has unprecedented advantages in capacity and access latency. Substantial innovation is required to fully exploit the potential benefits of integrating memory into processing units; current algorithms are unsuitable, because they are constrained by the von Neumann bottleneck. The PIs will design the Gigascale Analog Iterative Network Solver (GAINS), a system architecture to enable efficient in-situ data processing. GAINS alters the application-, architecture-, and logic/circuit-level abstractions that enable designers and developers at each layer to work independently. (i) It promotes matrices to a first-class data type. (ii) It integrates computataion and memory, avoiding pitfalls of conventional memory hierarchies. (iii) It exploits multi-valued representations in storage and computation. (iv) It replaces binary logic circuits with multi-valued analog circuits, reducing area overhead and power consumption. The PIs will investigate the effects of this changed paradigm on the design of algorithms in numerical optimization and machine learning.