Graduate Seminar: Energy-Efficient Memory Architectures for Future Datacenters and HPC Systems.
Location: McBryde 655
Speaker: Xun Jian, University of Illinois at Urbana-Champaign
Due to the large amount of power that datacenter and HPC systems already consume, energy-efficiency improvements are critical for sustainable performance scaling of these systems. Memories pose a looming power bottleneck for future data centers and HPCs due to both application trends and micro-architectural trends. Application trends such as in-memory big data computing and server virtualization are driving up memory capacity demands, while the micro-architectural trend of increasing core count per processor is driving up memory bandwidth requirement. For high memory capacity systems and high memory bandwidth systems, memory power can be the top source of overall system power.
In this talk, I discuss several unique power problems in the context of high capacity memory systems and high bandwidth memory systems, as well as the solutions to these problems. In the first part of the talk, I discuss why high memory capacity systems today pay a very high energy overhead today - 60% to 100% - to ensure reliable memory operations; I also describe how to significantly reduce this overhead via common-case optimized memory fault-tolerance while achieving the same memory redundancy and reliability requirements. In the second part of the talk, I discuss the unique power management challenges for memory networks, which enable scalable-capacity high bandwidth memory systems; I also describe a novel distributed power management solution for memory networks to effectively address these challenges.