A Scalable Approach to Modeling on Accelerated Neuromorphic Hardware.

Area of Science:

• Computational Neuroscience
• Hardware-Accelerated Computing
• Neuromorphic Engineering

Background:

• Neuromorphic systems offer expanded computational research possibilities but face challenges in balancing efficiency and usability.
• The BrainScaleS-2 system is a hybrid accelerated neuromorphic hardware architecture utilizing physical modeling.

Purpose of the Study:

• To present the software aspects of the BrainScaleS-2 system, focusing on enhancing its efficiency and usability.
• To introduce key components of the BrainScaleS-2 Operating System, including experiment workflow, API layering, software design, and platform operation.
• To derive software requirements through use cases and showcase their implementation.

Main Methods:

• Detailed description of the BrainScaleS-2 Operating System's architecture and design principles.
• Implementation of novel software features: multi-compartmental neurons, fast re-configuration for hardware-in-the-loop training, embedded processor applications, non-spiking operation mode, and interactive platform access.
• Development of sustainable hardware/software co-development strategies.

Main Results:

• Successful implementation of the BrainScaleS-2 Operating System, addressing efficiency and usability challenges.
• Demonstration of novel features enabling advanced neuromorphic computations and hardware-in-the-loop training.
• Establishment of a framework for interactive platform access and collaborative development.

Conclusions:

• The BrainScaleS-2 software advancements significantly improve the usability and efficiency of neuromorphic hardware.
• Future developments will focus on hardware scale-up, further enhancing system usability, and optimizing computational efficiency.
• The presented approach facilitates sustainable hardware/software co-development for next-generation neuromorphic systems.