Approximate Computing: Application Analysis and Hardware Design

It is normal to expect the processor to produce a correct result. But it may be the case that a software application does not need ”strict” correctness, but rather limited correctness. The goal of this project is to search for potential definition of ”non-strict” correctness, as well as investigate the improvements in performance and power consumption that can be achieved in this new model. This project proposes a hardware/software system design for applying the approximate computing idea to the wellknown problem of memory latency due to cache misses. We use three techniques to predict the values of cache misses: Last Value, Stride and Adaptive prediction. Because our approach is recovery-free, evaluation of quality of service after the execution is our main way to control the correctness of speculative predictions. Our evaluations show that the system we propose can successfully cover significant amount of L1/L2 cache misses. On average we can cover 62% of L1 misses with more aggressive, but less precise techniques, and 40% with the more precise, but less aggressive Adaptive technique. In a simple in-order model this corresponds to an average speedup of 12.5% and 10% respectively. For applu benchmark we can successfully cover more than 56% of L2 misses, that leads to 69% performance improvement.