BrainScaleS-2 — Heidelberg University

The BrainScaleS-2 accelerated neuromorphic system is an integrated circuit architecture for emulating biologically-inspired spiking neural networks. It was developed by researchers at the Heidelberg University and collaborators. Key features of the BrainScaleS-2 system include:

System Architecture

• Single-chip ASIC integrating a custom analog core with 512 neuron circuits, 131k plastic synapses, analog parameter storage, embedded processors for digital control and plasticity, and an event routing network
• Processor cores run a software stack with a C++ compiler and support hybrid spiking and non-spiking neural network execution
• Capable as a unit of scale for larger multi-chip or wafer-scale systems

Neural and Synapse Circuits

• Implements the Adaptive Exponential Integrate-and-Fire (AdEx) neuron model with individually configurable model parameters
• Supports advanced neuron features like multi-compartments and structured neurons
• On-chip synapse correlation and plasticity measurement enable programmable spike-timing dependent plasticity

Hybrid Plasticity Processing

• Digital control processors allow flexible implementation of plasticity rules bridging multiple timescales
• Massively parallel readout of analog observables enables gradient-based and surrogate gradient optimization approaches

Applications and Experiments

• Accelerated (1,000-fold compared to biological real time) emulation of complex spiking neural network dynamics, including configurable multi-compartmental cell morphologies
• Exploration of synaptic plasticity models and critical network dynamics at biological timescales
• Training of deep spiking neural networks using surrogate and exact gradient techniques
• Non-spiking neural network execution leveraging synaptic crossbar for analog matrix multiplication

The accelerated operation and flexible architecture facilitate applications in computational neuroscience research and novel machine learning approaches. The system design serves as a scalable basis for future large-scale neuromorphic computing platforms.

Related publications