Self-compensating accelerators for efficient approximate computing

摘要: Abstract Since most applications amenable to approximate computing involve large data paths, it is essential optimize accelerators, in addition designing individual arithmetic modules. To this end, we need minimize the error propagated through different With motivation, propose Self-Compensating Accelerators (SeCAs), that are constructed by combining modules such a way approximation completely or partially canceled within accelerator path, and thus cumulative at output reduced. For illustration purposes, use block-based adders recursive multipliers. Simulation results show proposed SeCAs help achieve significant benefits accuracy, while keeping other performance measures, i.e., speed, area power, unaffected. This quality gain can be exploited two ways: (1) employing when high cannot afforded, yet area/performance/power efficiency required; (2) using more aggressive approximations even further increase.