Unguided Machine Learning-Based Computation Offloading for Near-Memory Processing
Satanu Maity, Manojit Ghose, Avinash Kumar, Anol Chakraborty, Ankit Chakraborty · 2025
Recent data-intensive applications encounter memory wall bottlenecks in the traditional processor-centric computing architecture due to the need for frequent and extensive off-chip data movement. The emerging 3D-enabled near-memory processing (NMP) is a potential solution where computation can be offloaded and performed near memory, eliminating the need for expansive off-chip data transfer. Recently, some crucial works in this context have been proposed that use a supervised machine learning approach that requires a well-labeled dataset, which necessitates creating a time-consuming simulation process. Furthermore, their major focus was on profiling and offloading the entire application to the NMP side, causing a sub-optimal performance because of the low processing power on the NMP cores. This paper presents an unguided machine-learning-based approach (UCO) for computation offloading in NMP architecture that offloads only the memory-intensive regions of an application. The approach extracts several application performance characteristics and creates two clusters using them - one for the CPU side and another for the NMP side. The proposed approach's effectiveness is demonstrated through a series of comprehensive experiments with standard simulators and dataintensive benchmark applications. The results indicate that UCO improves performance (1.31x and 1.10x), saves energy (16% and 12%), and reduces off-chip data transfer (78% and 73%) compared to conventional and state-of-the-art approaches.