Approximate computing design exploration through data lifetime metrics

When designing an approximate computing system, the selection of the resources to modify is key. It is important that the error introduced in the system remains reasonable, but the size of the design exploration space can make this extremely difficult. In this paper, we propose to exploit a new metric for this selection: data lifetime. The concept comes from the field of reliability, where it can guide selective hardening: the more often a resource handles “live” data, the more critical it becomes, the more important it will be to protect it. In this paper, we propose to use this same metric in a new way: identify the less critical resources as approximation targets in order to minimize the impact on the global system behavior and therefore decrease the impact of approximation while increasing gains on other criteria.