Best is the Enemy of Good: Design Techniques for Low Power Tunable Approximate Application Specific Integrated Chips Targeting Media-Based Applications

With the possible end of the Moore's Law at the horizon, Approximate Computing has gathered momentum over the past years as a possible alternative for low power designs. Approximate computing trades better energy performance for tolerable inaccuracies at the output of the design. In this paper, we propose an application independent automated flow that converts a given design into an approximate version using either voltage scaling or power gating based techniques. The proposed model is shown to be effective for designing low power media type IPs (Intellectual Properties) based ASICs (Application Specific Integrated Chips). The model encompasses various automated techniques to identify logic within a given design which in turn can be leveraged for approximation. Following this identification the model uses a series of physical optimizations which lead to a tunable approximate circuit capable of operating in both approximate and accurate modes of operations depending on the environment and user constraints. The flow has been demonstrate to provide up to 35% power reduction in ASICs (operating in approximate mode) that implement imaging applications such as video decoding and image restoration IPs.