Arithmetic systems for low-power signal processing
M.J. Irwin, John R. Sacha · 1998
Digital signal processing has come to play an ever increasing role in computation, due to the emergence of such applications as multimedia and portable communication devices. In turn, circuit power dissipation has become a major concern in the design of high-performance VLSI-based systems, due to cooling considerations and extending battery life. An important aspect of both digital signal processing and low power VLSI design involves selecting appropriate number representations. Although the central role played by computer arithmetic in signal processing has led to the development of several specialized number representations to facilitate high speed operations, past low power research has tended to focus upon conventional fixed-point formulations. The present research examines some of the power characteristics of less common techniques, including floating point, logarithmic, and residue representations. A speech coding application, it is shown that the use of a floating point or logarithmic representation to trade precision for dynamic range can save power. Logarithmic representation is also attractive in that it replaces expensive fixed-point multiplies with less costly additions. Distributed arithmetic is a method for computing inner products which exchanges expensive multiplies for table lookups. A new method is presented for further savings: recoding the two's complement inputs into a nonredundant signed-digit representation eliminates much of the switching activity associated with the sign extensions in the higher-order bits.