Exploiting quality-energy tradeoffs with arbitrary quantization: special session paper

Abstract

Approximate computing aims to expose and exploit quality vs. efficiency tradeoffs to enable ever-more demanding applications on energy-constrained devices such as smartphones, or IoT devices. This paper makes the case for arbitrary quantization as a compelling approximation technique that exposes quality vs. energy tradeoffs and provides practical error guarantees. We present QAPPA (Quality Autotuner for Precision Programmable Accelerators), an autotuning framework for C/C++ programs that automatically minimizes the precision of each arithmetic and memory operation to meet user defined application level quality guarantees. QAPPA integrates energy models of precision scaling mechanisms to produce bandwidth and energy savings estimates for precision scalable accelerator designs. We show that QAPPA can exploit precision scaling mechanisms to meet arbitrary user-provided quality targets on the PERFECT benchmark suite to achieve significant energy savings and memory bandwidth reduction.