Self-compensating accelerators for efficient approximate computing

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