SpiNNaker

SpiNNaker. SpiNNaker (Spiking Neural Network Architecture) is a massively parallel neuromorphic engineering computing platform designed to simulate large-scale models of spiking neural networks in biological real time. Developed primarily at the University of Manchester as part of the Human Brain Project, it represents a radical departure from conventional von Neumann architecture by employing a unique interconnection network optimized for the communication patterns of neural simulations. The system's design enables the real-time modeling of up to a billion neurons, facilitating research in computational neuroscience, robotics, and machine learning.

Overview

The core objective of the SpiNNaker project is to provide a tool for modeling the human brain at an unprecedented scale, bridging the gap between biological neural networks and artificial intelligence. It operates as a manycore processor system where each chip integrates multiple ARM-based processing cores alongside a bespoke network-on-chip for efficient packet switching communication. Unlike traditional supercomputers like those at CERN or the National Science Foundation's facilities, SpiNNaker is optimized for the asynchronous, event-driven communication inherent to neural systems. This architecture supports the International Brain Laboratory and other global neuroscience initiatives, providing a platform for testing theories about brain function and neural coding.

Hardware architecture

The fundamental building block is the SpiNNaker chip, a system-on-chip incorporating eighteen ARM968 processor cores, with one typically reserved for system functions and the others for application computation. Each core is associated with local Tightly Coupled Memory and the system employs a multicast packet-switched network inspired by the Connection Machine and other massively parallel designs. The interconnection network uses a triangular toroidal topology to link tens of thousands of chips across multiple printed circuit boards housed in custom 19-inch rack enclosures. This design minimizes communication latency for neural spike events, a critical requirement for real-time simulation. Key collaborators in chip fabrication and design have included ARM Holdings and various partners within the European Union's research framework programmes.

Software and programming model

Programming SpiNNaker is facilitated by a software stack centered around PyNN, a simulator-independent application programming interface for neural network models, supported by the SpiNNaker Manchester team. The core runtime, written in C, runs a lightweight GNU/Linux kernel on the monitor core, while application cores execute bare-metal code for maximum efficiency. Neural models, defined using frameworks like PyNN or Brian (simulator), are compiled via the SpiNNTools toolchain into executable machine code and mapped onto the physical hardware. The SpiNNaker Application Runtime Kernel manages the low-level communication and real-time scheduling of neural processes, abstracting the complex hardware for researchers at institutions like the École Polytechnique Fédérale de Lausanne and the Allen Institute for Brain Science.

Applications and research

SpiNNaker has been deployed for a wide array of projects within the Human Brain Project and beyond, including real-time simulation of basal ganglia circuits relevant to Parkinson's disease, large-scale models of the cerebral cortex, and closed-loop systems for biorobotics. It has been used to control adaptive behaviors in robots like those developed at the University of Sussex and the Technische Universität München, interfacing with sensors such as Dynamic Vision Sensors. The platform also explores novel machine learning paradigms, including spiking neural network-based approaches to pattern recognition and sensory processing, providing a hardware testbed for algorithms that differ from those run on Google's Tensor Processing Unit or NVIDIA GPUs.

Development and project history

The SpiNNaker concept originated from research led by Steve Furber and the Advanced Processor Technologies Group at the University of Manchester in the early 2000s, with initial funding from the Engineering and Physical Sciences Research Council. Major development was subsequently propelled by a flagship grant from the European Research Council and its central role in the Human Brain Project, a large-scale scientific initiative funded by the European Commission. The first operational chips were fabricated around 2011, with the million-core machine, housed at the University of Manchester, completed in 2018. The project's legacy continues through its successor, SpiNNaker2, under development at the Technische Universität Dresden and other partners, aiming for greater energy efficiency and computational density.

Category:Supercomputers Category:Neuromorphic engineering Category:University of Manchester Category:Computer-related introductions in the 21st century