MamMAP: Efficient Memory Access Prediction with State-Space Models

Abstract

The effectiveness of machine learning based memory access prediction is currently limited by a fundamental trade-off: Transformers offer superior modeling of long-range dependencies but suffer from quadratic complexity, while Recurrent Neural Networks (RNNs) provide efficient inference but struggle with deep context retention. This paper introduces MamMAP, a novel prediction framework that leverages Mamba, a Selective State-Space Model architecture, to resolve this conflict. By replacing the attention mechanism with a selective scan and depth-wise causal convolution, MamMAP combines Transformers’ training parallelism with the linear-time inference efficiency of RNNs. To handle the stochastic nature of memory traces, we further propose a curriculum training strategy that utilizes variable-length history windows, enabling the model to robustly generalize across diverse program phases. We evaluate MamMAP using memory-intensive benchmarks from the SPEC CPU2006 and SPEC CPU2017. The results show that MamMAP sets a new performance standard, achieving 94% top-5 prediction accuracy and 65%prefetch coverage. Crucially, it delivers the lowest inference latency among evaluated models at 0.0698 ms/sample, translating to a 46% average IPC improvement.

Keywords

memory access prediction / state-space models / hardware prefetch

RIGHTS & PERMISSIONS

The Author(s) 2026.

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.