Go to main content

School of Computer Science Intranet

APT research areas

Discover our main research areas

Split-datapath execution for low power

Matt Evans

Abstract

Since the introduction of smartcards in the mid-1980s, the technology has improved drastically with many current cards incorporating tens of kilobytes of storage and even high-performance 32-bit processors. Most smartcards rely on making electrical contact with a reader when they are accessed.

Contactless smartcard ICs are simpler because of power limitations; they are powered directly from a magnetic field with no physical connections and must operate on very low power levels. Applications and users demand ever-increasing processing power from such cards but current high-performance processors consume too much power for many low-energy applications.

A scheme is proposed that can avoid unnecessary work by slicing the datapath and inferring partial results, rather than evaluating them. Such a system is described; for each partition metadata is supplied to the execution unit which describes the operand and allows unnecessary operations to be avoided.

A software model of a 32-bit ARM processor with a partitioned datapath shows that, when real programs are executed, partial ALU operations can be avoided -- or replaced with simpler operations -- up to 70\% of the time. This suggests that a split datapath would be advantageous in saving energy in a processor. Possible architectures are examined, and a VLSI implementation of an asynchronous byte-serial split-datapath ALU is presented. Power and timing analysis was performed on the layout, and the split-datapath ALU is compared to a contemporary, asynchronous 32-bit ALU in terms of performance, power use, energy efficiency and silicon area.

The split-datapath ALU implementation, being a prototype, showed some inefficiencies but may still have promise in the future. Further optmisations to the implementation would improve its efficiency, however real savings are likely to appear with very wide execution widths (such as 64 or 128-bit), and other power-hungry circuits such as system buses.

The thesis is available by ftp as GZipped PostScript (620K), or PDF (2.4MB).