A Resilient Interface for Approximate Data Access
Jo�ão Fabrício Filho, Isaías Felzmann, Rodolfo Azevedo, Lucas Wanner · 2019
Approximate memories offer lower energy cost while introducing errors to applications. If such errors affect essential parts of the application, the execution may fail, decreasing outputs quality and energy savings. We present AxRAM, a memory architecture interface for approximate data that allows applications to benefit from energy savings provided by approximate memories, while improving quality of results and failure rates. AxRAM places critical application data into non-approximate memory regions and implements a resilient addressing scheme that reduces invalid data accesses that lead to application crashes. On a voltage-overscaled SRAM, AxRAM reduces 50.97% of application crashes on average when compared to an unprotected approximate memory. Compared to a system using non-approximate SRAM, our proposal reduces energy consumption by 45.36% on average while providing at least 80% output quality.