Integrating Processing-In-Memory into HW/SW Co-Design for Automotive Embedded Systems

Continuous expansion of using intelligent learning workloads in modern vehicles, leads Electronic Control Units (ECUs) to manage massive volumes of data while dealing with hard real-time constraints. That said, ECUs must operate under strict power budget due to limited battery capacity and other safety and functional requirements. We study in this research the feasibility of integrating Processing-In-Memory (PIM) approach into the hardware/software co-design process for the emerging ECU architectures. The idea here is to allocate data-centric tasks to in-memory compute units so that computation occurs directly where data is stored. The proposed approach will allow avoiding expensive data traffic, which will improve processing performance and reduce energy consumption. Our work introduces a conceptual framework that reconciles the PIM strengths with automotive requirements and introduces techniques from dynamic voltage scaling to smarter memory management, and is limited to a conceptual architectural proposal without experimental validation or quantitative evaluation. We end up discussing some crucial opportunities and open challenges arising from the implementation of PIM in next-generation automotive systems.