BrainScaleS-2: Accelerated Analog Neuromorphic Computing Architecture

Technology Node

65 nm

Process Technology

Acceleration Factor

1000×

vs Biological Time

Neuron Support

Complex

Non-linear Dendrites

1. Introduction

The BrainScaleS (BSS) architecture represents a significant advancement in neuromorphic computing, combining analog physical model implementations of neurons and synapses with digital processing cores. The second-generation BrainScaleS-2 system, developed as part of the European Human Brain Project, marks a substantial improvement over its predecessor through the adoption of 65 nm technology and the integration of dedicated digital plasticity processing units.

2. BrainScaleS Architecture Overview

2.1 Analog Neural Core

The analog core implements continuous-time physical models of neurons and synapses, providing highly accelerated emulation of biological neural networks. The system operates with time constants several orders of magnitude smaller than biological systems, enabling rapid simulation of neural dynamics.

2.2 Digital Plasticity Processing

A key innovation in BSS-2 is the integration of a digital plasticity processing unit—a highly-parallel microprocessor specifically designed for learning operations in accelerated analog neuromorphic systems. This unit handles structural and parameter changes that occur on slower time scales compared to the analog neural dynamics.

2.3 System-on-Chip Design

The architecture features a neuromorphic System-on-Chip (SoC) comprising multiple digital CPU cores with specialized vector units connected through a network-on-chip. This design prioritizes event data while maintaining common address space for neurons and CPUs.

3. Technical Implementation

3.1 HICANN-X ASIC

The HICANN-X Application Specific Integrated Circuit represents the latest in-silico realization of the BSS-2 architecture. Built in 65 nm technology, it enables the integration of complex digital processing alongside analog neural circuits.

3.2 Neuron and Synapse Models

The system supports sophisticated neuron models including programmable ion channel emulation and inter-compartmental conductances. This enables modeling of non-linear dendrites, back-propagating action potentials, NMDA, and Calcium plateau potentials. The membrane dynamics can be described by:

$C_m \\frac{dV_m}{dt} = -g_L(V_m - E_L) - \\sum_i g_i(t)(V_m - E_i) + I_{ext}$

3.3 Calibration Framework

A custom software toolbox facilitates complex calibrated Monte-Carlo simulations, addressing the challenge of process variations in analog circuits. This calibration is essential for successful training and reliable operation.

4. Experimental Results

The BrainScaleS-2 system demonstrates significant improvements over the first generation. The integration of digital plasticity processing enables more flexible learning rules beyond basic STDP. The analog accelerator also supports vector-matrix multiplication, allowing both inference of deep convolutional networks and local learning with spiking neurons within the same substrate.

Figure 1: BrainScaleS Architecture Components

The architecture diagram shows the wafer-scale integration, BSS-1 ASIC, BSS-2 neuron design, and exemplary membrane voltage traces demonstrating the system's capability to emulate complex neural dynamics.

Figure 2: Neuromorphic SoC Architecture

The SoC architecture illustrates multiple processor tiles with vector units and analog cores connected via high-bandwidth links and network-on-chip, featuring specialized function tiles for memory control and SERDES I/O.

5. Code Implementation

The system utilizes PyNN, a simulator-agnostic neural network description language, providing a unified software interface. Below is a simplified example of neuron configuration:

# PyNN code example for BrainScaleS-2
import pyNN.brainscales as bss

# Configure neuron parameters
neuron_parameters = {
    'tau_m': 10.0,      # membrane time constant
    'cm': 1.0,          # membrane capacitance
    'v_rest': -70.0,    # resting potential
    'v_thresh': -55.0,  # threshold potential
    'tau_syn_E': 5.0,   # excitatory synapse time constant
    'tau_syn_I': 5.0    # inhibitory synapse time constant
}

# Create neuron population
population = bss.Population(100, bss.IF_cond_exp, neuron_parameters)

# Configure plasticity rule
stdp_model = bss.STDPMechanism(
    timing_dependence=bss.SpikePairRule(),
    weight_dependence=bss.AdditiveWeightDependence()
)

6. Future Applications

The BrainScaleS-2 architecture opens new possibilities for neuromorphic computing applications. The combination of accelerated analog emulation with digital programmability makes it suitable for real-time AI systems, brain-inspired computing research, and low-power edge AI applications. Future developments may focus on scaling to larger neural networks, improving energy efficiency, and enhancing the programmability of learning rules.

Original Analysis

The BrainScaleS-2 architecture represents a sophisticated approach to neuromorphic computing that bridges the gap between biological plausibility and computational efficiency. By combining analog physical models with digital programmability, it addresses fundamental challenges in neuromorphic hardware design. The system's 1000× acceleration factor compared to biological time scales enables practical research applications that would otherwise require impractically long simulation times.

Compared to other neuromorphic approaches like IBM's TrueNorth and Intel's Loihi, BrainScaleS-2 offers unique advantages in biological realism through its analog implementation. While digital systems like Loihi provide greater programmability, the analog approach of BrainScaleS-2 potentially offers better energy efficiency for certain classes of neural computations. This aligns with trends observed in recent neuromorphic research, where hybrid analog-digital approaches are gaining traction for their balanced performance characteristics.

The integration of a dedicated digital plasticity processor addresses a key limitation of purely analog systems: the difficulty of implementing complex, programmable learning rules. This innovation enables BrainScaleS-2 to support not only fixed STDP but also more sophisticated learning mechanisms, making it more versatile for research into neural plasticity and learning algorithms.

The system's support for both spiking neural networks and deep learning inference through vector-matrix multiplication demonstrates a pragmatic approach to the current AI landscape. This dual capability allows researchers to explore brain-inspired computing while maintaining compatibility with mainstream deep learning approaches. The calibration framework for managing analog process variations shows sophisticated engineering that acknowledges and addresses the practical challenges of analog neuromorphic computing.

Looking forward, architectures like BrainScaleS-2 could play a crucial role in developing more energy-efficient AI systems, particularly for edge computing applications where power constraints are critical. The European Human Brain Project's continued investment in this technology underscores its potential significance for both neuroscience research and practical AI applications.

7. References

Schemmel, J., et al. "A Wafer-Scale Neuromorphic Hardware System for Large-Scale Neural Modeling." ISCAS 2010.
Indiveri, G., et al. "Neuromorphic silicon neuron circuits." Frontiers in Neuroscience, 2011.
Davies, M., et al. "Loihi: A Neuromorphic Manycore Processor with On-Chip Learning." IEEE Micro, 2018.
Merolla, P. A., et al. "A million spiking-neuron integrated circuit with a scalable communication network and interface." Science, 2014.
Pei, J., et al. "Towards artificial general intelligence with hybrid Tianjic chip architecture." Nature, 2019.
European Human Brain Project. "Neuromorphic Computing Platform." https://www.humanbrainproject.eu
IEEE Spectrum. "The Quest for Artificial Intelligence that Mimics the Brain." https://spectrum.ieee.org

Table of Contents