Approximation through logic isolation for the design of quality configurable circuits

Intrinsic application resilience, a property exhibited by many emerging application domains, allows designers to optimize computing platforms by approximating selected computations within an application without any perceivable loss in its output quality. At the circuit level, this is often achieved by designing circuits that are more efficient but realize slightly modified functionality. Most prior efforts on approximate circuit design hardwire the degree of approximation into the implementation. This severely limits their applicability, as intrinsic resilience significantly varies both across and within applications, and often the same computation needs to be executed at different levels of accuracy when the application processes a different input or used in a different context. To address this limitation, in this work, we propose a new approach to design quality configurable circuits that are equipped to modulate their output accuracy and energy at runtime. Our approach, approximation through logic isolation, identifies portions of logic in the circuit that consume significant power, but contribute only minimally to output accuracy. One or more approximate modes of circuit operation are then enabled by isolating the identified logic (using muxes, latches or power gating cells) to benefit power while satisfying the desired output accuracy. We propose a systematic methodology to transform a given circuit into a quality-configurable circuit by applying the proposed technique. Our methodology generates a favorable energy-quality trade-off by deliberately creating opportunities for error compensation between multiple logic islands that are simultaneously isolated. This enables more aggressive approximation for a given output quality, leading to a superior power benefits. We evaluate the proposed methodology using a wide range of arithmetic circuits, complex modules and datapaths. The synthesized quality configurable circuits support 3 quality modes viz. accurate, <;0.1% average error, and <;0.2% average error. Power improvements achieved in the approximate modes are 8.4%-34.5% and 17%-51.5%, respectively.