Energy-Efficient Design of the Reorder Buffer

摘要

Some of today's superscalar processors, such as the Intel Pentium III, implement physical registers using the Reorder Buffer (ROB) slots. As much as 27% of the total CPU power is expended within the ROB in such designs, making the ROB a dominant source of power dissipation within the processor. This paper proposes three relatively independent techniques for the ROB power reduction with no or minimal impact on the performance. These techniques are: 1) dynamic ROB resizing; 2) the use of low-power comparators that dissipate energy mainly on a full match of the comparands and, 3) the use of zero-byte encoding. We validate our results by executing the complete suite of SPEC 95 benchmarks on a true cycle-by-cycle hardware-level simulator and using SPICE measurements for actual layouts of the ROB in 0.5 micron CMOS process. The total power savings achieved within the ROB using our approaches are in excess of 76% with the average performance penalty of less than 3%.