Abstract
We present a novel approach for power minimization of CMOS combinational circuits based on transforming power-expensive inputs to functionally equivalent power-optimal inputs. We assume that the environment information is available in the form of either word-level statistics or a long representative input sequence. From the environment information, we derive power-effective input transformations that will save power for any input stream that satisfies the environment properties. We synthesize a transform module that is appended at the primary inputs of the module. The proposed approach can be applied at the system-level by propagating environment information to the primary outputs of the system and then deriving transform circuits for each module. As an example of real-world application, we present results for DCT module in a JPEG compression chip with different input environments (eg. face recognition dataset, weather forecasting). Experimental results are highly encouraging.