SecPIMlet: Confidential Computing for Spatio-temporally Shared Processing-in-Memory

Sunho Lee, Minwoo Noh, Jeongwon Choi, Kwanghoon Choi, Igjae Kim, Chang Hyun Park, and Jaehyuk Huh, "SecPIMlet: Confidential Computing for Spatio-temporally Shared Processing-in-Memory", the 59th International Symposium on Microarchitecture ( MICRO ), October 2026

Paper Slide

As memory-intensive applications, particularly language models, continue to grow, Processing-in-Memory (PIM) has emerged as a promising architecture to alleviate memory bandwidth limitations. We find that PIM systems need to support both spatio and temporal sharing between domains to extract the maximal performance. However, current PIM systems lack the necessary protection to securely support such sharing. To address these issues, we introduce SecPIMlet, a trusted multi-domain PIM that enables efficient spatio-temporal sharing. For fast and confidential spatio-temporal sharing, SecPIMlet protects the PIM-resident data and execution context and also protects in-transit data between the CPU and PIM. Our evaluation shows that SecPIMlet delivers 175.4% of throughput improvement compared to prior inefficient spatio-temporal sharing, while providing a robust security solution.