This paper examines the effects of multithreaded pipelining on the CPI (cycles per instruction) of a RISC processor. The desired CPI in a conventional (single-threaded) RISC processor is one instruction per cycle. However, the CPI is typically more than one because of data hazards, control hazards, and resource hazards in the pipeline.
A multi-threaded processor performs a context switch between every instruction. Multi-threaded pipelining holds out the promise of achieving a lower CPI because it can eliminate data and control hazards, and mask the effects of memory latency. However, multi-threaded pipelining reduces cache hit ratios and requires more chip area to implement.
In this paper, we present a model for predicting the CPI of a multi-threaded pipelined processor. We also present the results of trace-driven simulations of single- and multi-threaded processors. These data show that, for reasonable implementation technologies, and taking into account the chip area penalty, a multi-threaded processor can achieve a lower CPI than a single-threaded processor.
Copyright 1991 by Engineering Institute of Canada.
Full text. BibTeX entry.