Complex Load-Value Predictors: Why We Need Not Bother

Memory accesses continue to represent a major performance bottleneck and much remains to be done to tolerate their latencies. A large body of work exists that presents load-value prediction as an effective means to hide some of the memory latency. To increase the prediction accuracy and hence the performance, researchers have proposed more complex and larger predictor designs. This paper re-evaluates load-value predictors and examines the tradeoffs between predictor size, latency, energy consumption and performance. We demonstrate that even though complex predictors deliver the highest accuracy, they are unlikely to be implemented in hardware because they provide a much worse energy-performance tradeoff than simpler predictors with moderate sizes.