Memory access algorithm for low energy CPU/GPU heterogeneous systems with hybrid DRAM/NVM memory architecture
Tsai-Kan Chien, Lih‐Yih Chiou, Chieh-Wen Cheng, Shyh-Shyuan Sheu, Peihua Wang, Ming‐Jinn Tsai, Chih‐I Wu · 2016
In modern mobile devices, the CPU and GPU can jointly access one shared memory, typically consisting of only DRAM, by using heterogeneous uniform memory access (hUMA). This paper proposes an access algorithm, migration by data access tendency (MiDATE), for a hybrid main memory architecture consisting of resistive random access memory and DRAM. The MiDATE algorithm can operate on this hybrid memory architecture, reducing energy consumption by 83% while maintaining approximately the same speed as that of DRAM-only systems. Using the MiDATE algorithm can, on average, reduce unfairness by 36.5% compared with that of conventional hUMA methods, increasing system speed.