Applying Perceptrons to Speculation in Computer Architecture- Neural Networks in Future Microprocessors

Modern microprocessors make use of speculation or predictions about futureprogram behavior to optimize the execution of programs. Perceptrons aresimple neural networks that can be highly useful in speculation for theirability to examine larger quantities of available data than more commonlyused approaches and identify which data lead to accurate results. This workfirst studies how perceptrons can be made to predict accurately when theydirectly replace the traditional pattern table predictor. Different training methodsperceptron topologies and interference reduction strategies areevaluated. Perceptrons are then applied to two speculative applications: datavalue prediction and dataflow critical path prediction. Several novel perceptron-based prediction strategies are proposed for each application that cantake advantage of a wider scope of past data in making predictions thanprevious predictors could. These predictors are evaluated against local tablebasedapproaches on a custom cycle-accurate processor simulator and areshown on average to have both superior accuracy and higher instruction-percycleperformance. This work is addressed to computer architects and computerengineering researchers.