An Updated Evaluation of ReCycle

Process variation reduces a pipeline’s maximum attainable frequency by creating unbalance in the stage delays. As a result, the pipeline ends up cycling with a period close to that of the slowest pipeline stage. ReCycle was proposed in ISCA 2007 as a framework for comprehensively applying cycle time stealing to balance the stage delays under process variation, thereby allowing the pipeline to cycle with a period close to the average latency of the stages. This paper duplicates the evaluation of ReCycle with more realistic pipeline and critical path models than in the original paper. Most notably, we do not assign one cycle to the feedback path of each pipeline loop. As a result, our pipeline contains five singlecycle loops. This is significant because these loops are not amenable to cycle time stealing. In this updated environment, ReCycle is able to regain on average only 40% of the performance lost to process variation. In contrast, in the original paper, ReCycle regained 64%. Moreover, we find that further adding Donor stages does not significantly increase performance. Consequently, we propose to extend the Donor algorithm by applying Forward Body Biasing (FBB) to single-cycle loops when they become critical. With ReCycle, Donor stages, and FBB, we regain on average 90% of the performance lost to variation — still short of the roughly 110% regained by ReCycle and Donor stages alone in the original ReCycle paper.