Abstract: Recently,the emerging non-volatile memory(NVM) exemplified by Phase Change Memory(PCM) has been considered to take the place of conventional DRAM in the processors due to their characteristics of large capacity and little static power．However,the overhead of long write latency can cause severe performance degradation．We evaluated the feasibility of PCM based GPU memory architecture．Based on the analysis of unique memory access behaviors captured from GPU benchmarks,a dedicated buffer was designed to alleviate the pressure of frequently accessing PCM．Simulation results prove the efficiency of our proposed dedicated buffer and show its great potential for PCM based GPU architecture.

Key words: Phase change memory(PCM),Graphics processing unit(GPU),Buffer,Memory hierarchy