Performance Implications of Processing-in-Memory Designs on Data-Intensive Applications

摘要

The popularity of data-intensive applications and recent hardware developments drive the re-emergence of processing-in-memory (PIM) after earlier explorations several decades ago. To introduce PIM into a system, we must answer a fundamental question: what computation logic should be included into PIM? In terms of computation complexity, PIM can be either relatively simple, fixedfunctional, or fully programmable. The choice of fixedfunctional PIM and programmable PIM has direct impact on performance. In this paper, we explore the performance implications of fixed-functional PIM and programmable PIM on three data-intensive benchmarks-including a real data-intensive application. Our results show that - with PIMs - we obtain 2.09x-91.4x speedup over no PIM cases. However, the fixed-functional PIM and programmable PIM perform differently across applications (with performance difference up to 90%). Our results show that neither fixed-functional PIM nor programmable PIM can perform optimally in all cases. We must decide the usage of PIM based on the characteristics of the workload and PIM (e.g., instruction-level parallelism), and the PIM overhead (e.g., PIM initialization and synchronization overhead).