Edge AI deployments are becoming increasingly complex, necessitating energy-efficient solutions for resource-constrained embedded systems. Approximate computing, which allows for controlled inaccuracies in computations, is emerging as a promising approach for improving power and energy efficiency. Among the key techniques in approximate computing are approximate arithmetic operators (AxOs), which enable application-specific optimizations beyond traditional computer arithmetic hardware reduction-based methods, such as quantization and precision scaling. Existing design space exploration (DSE) frameworks for approximate computing limit themselves to selection-based approaches or custom synthesis at fixed abstraction levels, which restricts the flexibility required for finding application-specific optimal solutions. Further, the tools available for the DSE of AxOs are quite limited in terms of exploring different approximation models and extending the analysis to different granularities. To this end, we propose AxOSyn, an open-source framework for the DSE of AxOs that supports both selection and synthesis approaches at various abstraction levels. AxOSyn allows researchers to integrate custom methods for evaluating approximations and facilitates DSE at both the operator-level and application-specific. Our framework provides an effective methodology for achieving energy-efficient, approximate operators.
翻译:暂无翻译