A Critical Assessment of DRAM-PIM Architectures - Trends, Challenges and Solutions

Recently, we are witnessing a surge in DRAM-based Processing in Memory (PIM) publications from academia and industry. The architectures and design techniques proposed in these publications vary largely, ranging from integration of computation units in the DRAM IO region (i.e., without modifying DRAM core circuits) to modifying the highly optimized DRAM sub-arrays inside the banks for computation operations. Additionally, the underlying memory type, e.g., DDR4, LPDDR4, GDDR6 and HBM2, for DRAM-PIM is also different. This paper presents the assessment of DRAM-PIM architectural design decisions adapted in all DRAM-PIM publications. Our study presents an in-depth analysis of computation unit placement location, i.e., from the chip-level down to DRAM sub-array-level, and discusses the implementation challenges for a computation unit in various regions of commodity DRAM architectures. We also elaborate on the architectural bottlenecks associated with the scalability of DRAM-PIM performance and energy gains, and present architectural approaches to address the issues. Finally, our assessment covers other important design dimensions, such as computation data formats and DRAM-PIM memory controller design.