Lecture Notes in Computer Science, 2005, Volume 3471/2005, 30-45, DOI: 10.1007/11574859_3

Bit-Sliced Datapath for Energy-Efficient High Performance Microprocessors

Sumeet Kumar, Prateek Pujara and Aneesh Aggarwal

View Related Documents

Abstract

In the recent years, both power and performance have become important in the design of microprocessors. In this paper, we investigate exploiting the small-sized data values for energy-efficient high performance microprocessors. For this purpose, we bit-slice the execution core (which includes the functional units, register files, data caches, and forwarding logic), so that small portions of the data are operated upon in different bit-slices. The bit-slices operating upon the higher order bits are activated only if required, saving significant energy consumption. We also investigate further advantages facilitated by bit-slicing such as energy savings obtained by reducing the number of ports provided in the higher order bit-slices and by “shutting off” bit-slices to reduce leakage energy consumption. We found that a significant energy saving can be obtained in the register file (about 20%) and the Level-1 Data Cache (about 30%) with a negligible loss of only about 2% in the instruction throughput. Our studies also showed almost 20% savings in the register file leakage energy consumption, when the unwanted higher order bit-slices are “turned off”. Bit-slicing also helps in reducing the latency of the different hardware structures, which can facilitate clock speed improvements.
An erratum to this chapter can be found at http://dx.doi.org/10.1007/11574859_13 .

Fulltext Preview

Image of the first page of the fulltext document