processing in memory

**Processing-in-Memory (PIM)** is **the execution of computation directly within or adjacent to memory arrays rather than shuttling data between separated memory and processor components**—a fundamental architectural shift to eliminate the memory wall bottleneck that dominates power and latency in modern systems. **Core PIM Technologies:** - HBM-PIM: Samsung's approach embeds logic layers in HBM stacks (compute within 3D memory cube) - UPMEM: prefetch processing near DRAM arrays with a lightweight ISA - AiM (AI in Memory): analog in-memory computing for neural networks - DRAM with embedded compute: transistors directly accessible to DRAM cells **Memory Architecture Considerations:** - Eliminates repeated memory-processor round-trips (critical for bandwidth-bound ML inference) - DRAM HBM2 PIM adds a logic layer beneath memory stacks for near-DRAM computation - Near-data processing (NDP) vs true in-memory compute represents spectrum of solutions - PIM ISA design: limited instruction set for domain-specific operations **Applications and Programming Challenges:** - Database query acceleration (WHERE filtering near storage) - ML inference kernels (matrix multiply in DRAM) - Data analytics (aggregation, reduction operations) - Programming model complexity: how to express PIM-compatible code in standard frameworks - Data layout optimization: tiling for memory hierarchy still critical **Impact and Future:** PIM promises orders-of-magnitude improvements in memory bandwidth utilization and energy efficiency for data-intensive workloads, though adoption requires rethinking compiler toolchains and algorithmic approaches to fully realize memory-compute fusion benefits.