QAPPA: A Framework for Navigating Quality-Energy Tradeoffs with Arbitrary Quantization

Quantization naturally exposes knobs in hardware to trade fidelity for efficiency: the more bits that are used to represent the data, the higher the storage and computation overheads. With the emergence of approximate computing research, we set out to answer the following question: how effective is quantization in trading off quality for efficiency, and how does it compare with other approximation techniques? This paper makes the case for quantization as a general approximation technique that exposes fine quality vs. energy tradeoffs and provides practical error guarantees. We assume arbitrary quantization levels, and focus on the hardware subsystems that are affected by quantization: memory and computation. We present QAPPA (Quantization Autotuner for Precision Programmable Accelerators), an autotuner for C/C++ programs that automatically tunes the precision of each arithmetic and memory operation to meet user defined application level quality guarantees. QAPPA integrates energy models of quantization scaling mechanisms to produce bandwidth and energy savings estimates for custom accelerator designs. We use the analysis produced by QAPPA to compare the effectiveness of arbitrary quantization against voltage overscaling and neural approximation. Our analysis shows that when using the right quantization scaling mechanisms in hardware, quantization provides significant energy efficiency benefits over voltage overscaling and comparable energy efficiency gains over neural approximation. Additionally, quantization offers more predictable error degradation and fully tunable error bounds.