memory centric computing architectures

**Memory-Centric Computing Architectures** — Design paradigms that place memory at the center of computation, minimizing data movement by bringing processing capabilities closer to where data resides. **Processing-In-Memory Approaches** — PIM architectures embed computational logic directly within memory chips, performing operations on data without transferring it to external processors. DRAM-based PIM adds simple ALUs to memory banks, enabling bulk bitwise operations and reductions at memory bandwidth speeds. SRAM-based PIM exploits the analog properties of memory arrays to perform multiply-accumulate operations for neural network inference. Hybrid approaches like Samsung's HBM-PIM integrate processing elements within high-bandwidth memory stacks, providing substantial bandwidth improvements for memory-bound workloads. **Near-Data Processing Architectures** — Near-data processing places compute units adjacent to memory or storage rather than inside the memory array itself. Smart SSDs with embedded FPGAs or ARM cores filter and preprocess data before sending results to the host, reducing PCIe bandwidth demands. Computational storage devices perform pattern matching, compression, and database scans at the storage layer. Active memory systems attach lightweight processors to each memory module, creating a distributed processing fabric that scales with memory capacity. **Programming Models and Challenges** — Memory-centric architectures require new programming abstractions that express data locality and in-situ operations. Compiler analysis must identify operations suitable for offloading to PIM units versus those requiring traditional processor execution. Data layout optimization becomes critical since PIM operations typically work on data within a single memory bank or row. Coherence between PIM-modified data and cached copies in the host processor requires careful protocol design to avoid stale reads and lost updates. **Application Domains and Performance Impact** — Graph analytics benefit enormously from PIM due to irregular memory access patterns that defeat caching. Database operations like selection, projection, and aggregation can execute entirely within memory, eliminating data transfer overhead. Genome sequence alignment performs character comparisons in bulk using bitwise PIM operations. Machine learning inference on edge devices uses analog PIM for energy-efficient matrix-vector multiplication. Studies show 10-100x energy reduction and 5-50x performance improvement for memory-bound workloads compared to conventional architectures. **Memory-centric computing architectures address the fundamental data movement bottleneck in modern systems, promising transformative improvements in performance and energy efficiency for data-intensive parallel workloads.**