A Processing-in-Memory Taxonomy and a Case for Studying Fixed-function PIM

The emergence of die-stacking technology with mixed logic and memory processes has brought about a renaissance in “processing in memory” (PIM) concepts, first envisioned decades ago. For some, the PIM concept conjures an image of a complete processing unit (e.g., CPU, GPU) integrated directly with memory, perhaps on a logic chip 3D-stacked under one or more memory chips. However, PIM potentially covers a very wide spectrum of compute capabilities embedded in/with the memory. This position paper presents an initial taxonomy for in-memory computing, and advocates for the exploration of simpler computing mechanisms in the memory stack in addition to fully-programmable PIM architectures.