Rapid, low-power loop execution in a network of functional units

Abstract

The need for high-performance computing and low-power operation has led to the emergence of new processor architectures, with most recent designs based on the combination of multiple cores and multiple threads per core. In our work, we are exploring an architecture of multiple instruction pipelines, which merge into a common back-end, formed as a network of functional units. We focus on the back-end in this paper, and in particular, on a rapid, low-power execution of loops, based on data flow. We dispatch the loop body instructions on the network of functional units only once, and we then let the loop execute in a dataflow manner, without any other instruction issue before loop completion. In this way, we do not only speed up the loop execution but we also save energy, since during the execution of the loop the whole front end of the pipeline is not used and can be turned off. We have simulated the functional unit network on microarchitecture level, running a number of Livermore loops. The results we obtained show that the proposed architecture can accelerate loop execution by up to N/k, for a network of N units and loop body size of N instructions, and an issue rate of k instructions per cycle. Copyright © 2013 ACM.