ATHENA: A CONFIGURABLE, DYNAMIC AND AUTONOMOUS ACCELERATOR FOR GENERAL PURPOSE APPLICATIONS

Abstract

Reconfigurable architecture has been successfully used as general-purpose accelerators providing improvement in both performance and energy efficiency. However, these gains come at the cost of a significant area overhead due to the large number of functional units that are generally used in the CGRAs, which can make its integration in embedded systems with hard area constraints prohibitive. In order to reduce the area overhead of CGRAs, this work proposes ATHENA (A THin rEcoNfigurable Array). ATHENA dynamically maps the applications’ kernels to be executed on the CGRA. The mapping is done through dedicated hardware (dynamic configuration generator), which is implemented in a five-stage pipeline. ATHENA distributes its computation over time, allowing functional units to be reused by mapping operations to the same unit, but in different cycles. ATHENA is implemented in the gem5 simulator and uses the mibench benchmark to evaluate the proposed system. ATHENA was also synthesized using the cadence synthesis tool with 45 nm technology. The results show ATHENA was able to accelerate up to 1.72x while saving 37% in energy, on average, by adding only 2.4% of area overhead on an 8-wide superscalar processor. Thus, ATHENA is a promising solution for mobile and embedded devices providing performance and energy improvements at low area cost.