Performance and energy efficiency in simple simultaneous multithreading processor cores

[摘要] Simultaneous multithreading, where instructions from different threads share processor resources, has shown promise in delivering high throughput with little area and power overhead. We compare where in the performance energy-efficiency space alternative simple simultaneous multithreading configurations lie, leveraging standard industry tools to estimate area and power from high level hardware descriptions. We find sharing function units among threads can improve energy efficiency over duplicating the function unit set for each thread. A good choice for the number of threads sharing a function unit ensures the function unit is not overloaded. Sharing front-end pipeline logic does not improve performance or energy efficiency over either duplicating the full pipeline or just duplicating the front-end pipelines for each thread. Different arbitration policies for use of function units do not impact performance much, but they do have a large impact on the power of the core, so the simplest arbitration policy should be used to maximize energy efficiency. Operand bypassing, an obvious optimization for a pipeline which does not share function units, is not obviously better when function units are shared, improving performance at the cost of reduced energy efficiency.

[发布机构] Massachusetts Institute of Technology