PIM or CXL-PIM? Understanding Architectural Trade-offs Through Large-Scale Benchmarking

TL;DR

This paper compares processing-in-memory (PIM) and CXL-PIM architectures through large-scale benchmarking, revealing workload-dependent tradeoffs and regimes where each approach performs better, guiding future near-memory system design.

Contribution

It provides the first large-scale, real hardware comparison of PIM and CXL-PIM, analyzing performance tradeoffs across different workload regimes.

Findings

Unified address space in CXL-PIM reduces data transfer bottlenecks.

Performance advantages of PIM or CXL-PIM depend on workload phase and dataset size.

Practical guidance for designing near-memory systems based on workload characteristics.

Abstract

Processing-in-memory (PIM) reduces data movement by executing near memory, but our large-scale characterization on real PIM hardware shows that end-to-end performance is often limited by disjoint host and device address spaces that force explicit staging transfers. In contrast, CXL-PIM provides a unified address space and cache-coherent access at the cost of higher access latency. These opposing interface models create workload-dependent tradeoffs that are not captured by small-scale studies. This work presents a side-by-side, large-scale comparison of PIM and CXL-PIM using measurements from real PIM hardware and trace-driven CXL modeling. We identify when unified-address access amortizes link latency enough to overcome transfer bottlenecks, and when tightly coupled PIM remains preferable. Our results reveal phase- and dataset-size regimes in which the relative ranking between the two architectures reverses, offering practical guidance for future near-memory system design.