Platform-Based Design of Embedded Neuromorphic Systems

Neuromorphic systems are integrated circuits designed to mimic the event-driven computations in a mammalian brain. These systems enable the execution of machine learning applications that are designed using spiking neural networks (SNNs). To cope with the growing complexity of such systems, challenges in integrating emerging non-volatile memory (NVM) technologies, and faster time-to-market pressure, efficient design methodologies are needed. Here, we describe platform-based design, a core concept from the embedded systems world that is likely to address those design issues associated with embedded neuromorphic systems. In a platform-based design methodology, a hardware platform is abstracted from its system software, making the hardware and software developments orthogonal to allow a more effective exploration of alternative solutions. Platform-based design methodology facilitates the reuse of the system software for many different hardware platforms. We describe how platform-based design methodologies can be applied to neuromorphic system design. Specifically, we show that a given system software framework can be optimized to achieve performance, energy, and reliability goals of a target NVM-based neuromorphic system. We also show the changes necessary to port the system software framework from one hardware platform to another and across different neuromorphic architectures and manufacturing process/NVM technology alternatives.