Energy savings techniques in out-of-order pipeline through value approximation of instructions with data dependencies

Abstract

<p>Approximate computing has emerged as one of the areas studied over the past few years to improve the performance and energy consumption computers. Approximate computing tolerates imprecision during computation, and it produces data values that are close to the actual outputs obtained from exact computation. From software to circuit level, approximate computing techniques have been applied across all computing domains. This study was carried out on microarchitectural level, where dependencies between two instructions are relaxed in the scheduling unit. This research proposed a technique that allows dependent instructions to execute without waiting for values produced by their producer instructions. This process enabled schedulers to skip certain pipeline processes such operand rename lookup, and instruction wake-up in the instruction scheduler queue to provide additional energy savings. The results of this work revealed an average performance acceleration of 1.25x. In addition, the total of energy savings was achieved at 4.6% for approximation cases that produced tolerable error at the output.</p>