Probability-Based Memory Access Controller (PMAC) for Energy Reduction in High Performance Processors

The increasing transistor density due to Moore's law scaling continues to drive the improvement in processor core performance with each process generation. The additional transistors are used to widen the pipeline, increase the size of the out-of-order instruction scheduling window, register files, queues and other pipeline data structures to extract high levels of instruction level parallelism and improve upon single- threaded performance. Such dynamically scheduled superscalar processor cores speculatively fetch and execute several instructions far ahead in a program, along the program path predicted by its branch predictors. During branch mispredictions, the architectural state of high performance processor cores can be restored at cost of high latency penalties, but the speculative memory requests sent by data memory access instructions on the mispredicted paths cannot be revoked. Such memory requests alter the data arrangement across memory hierarchy and result in wasted memory transactions, bandwidth and energy consumption. Even with low branch misprediction rates, these processor cores spend significant time on mispredicted program paths. In this thesis, we propose a probability based memory access controller to curb the data memory requests sent along mispredicted paths and achieve energy and memory bandwidth savings with minimum impact on performance. It computes path probability of instructions and throttles memory access instructions with low probability of execution. A deterministic or dynamically varying probability value is used as a threshold to control speculative memory requests sent to the memory hierarchy. The proposed design with a dynamic threshold reduces up to 51% of wrong path memory accesses and maximum of 31% of wrong path execution while achieving power savings up to 9.5% and maximum of 6.3% improvement in IPC/Watt in a single core processor system