MEMPLEX: A Multi-Chiplet NUMA Architecture with Data Replication and Migration

Paper i proceeding, 2025

As the semiconductor industry struggles with the diminishing returns of Moore's law and explores innovative solutions for integrating more resources on a chip, multi-chiplet chips offer a cost-efficient alternative to large monolithic chips due to their higher yield. However, chiplet-based systems inherently exhibit Non-Uniform Memory Access (NUMA) characteristics and, therefore, suffer from slow remote accesses. Although data placement in multi-chiplet NUMA systems can be optimized in software, currently, there are no hardware mechanisms to dynamically improve data placement in DRAM distributed across chiplets. Our experiments show that this leads to wasting a significant fraction of system performance compared to a hypothetical system with ideal data placement. Our work addresses this problem by introducing MEMPLEX, a novel memory system for multi-chiplet NUMA architectures, which offers data replication and migration in the memory nodes of a multi-chiplet system. MEMPLEX allocates a small fraction of each memory node to construct a DRAM cache and offers their remaining capacity to a shared flat address space with hardware migration. In a nutshell, MEMPLEX DRAM cache attracts data of the working set to the local memory node and decides whether to migrate them upon eviction based on their usage in the cache. Thereby, MEMPLEX improves data locality, regains a large fraction of the above performance overhead, and offers substantial energy savings.