SpiNNaker
SpiNNaker (a contraction of Spiking Neural Network Architecture) is a million-core supercomputer designed at the University of Manchester to simulate the behaviour of a billion neurons. Conceived and led by Professor Steve Furber, one of the original designers of the ARM microprocessor, the project formally began in 2005 and the system’s first chips were delivered in 2009. The SpiNNaker system became fully operational in 2018, with around a million ARM processors simulating a billion neurons.
A key part of SpiNNaker’s funding came from the EU’s Human Brain Project (HBP), started by Henry Markram in 2013 with the ambitious goal of simulating an entire human brain within a decade. In 2015, scientists who felt they were not getting their share of the funding forced Markram out and the HBP lost momentum. Nevertheless, in 2019 HBP announced a contribution to the development of a second generation SpiNNaker chip, called SpiNNaker 2. The company that is commercialising this is called SpiNNcloud systems, and is based in Dresden.
IBM: TrueNorth and NorthPole
TrueNorth was launched by IBM in 2014. It comprised 4,096 cores, each with 256 programmable simulated neurons, giving a total of a million neurons. Each neuron had 256 programmable “synapses” to convey signals, so the system had a total of just over 268 million synapses.
TrueNorth’s cores combined memory and processors, thus avoiding one of the von Neumann architecture bottlenecks. IBM claimed the system’s power consumption of 70 milliwatts meant its power density was 1/10,000th that of conventional microprocessors. It also operated without a governing system clock.
In 2023, IBM released the NorthPole chip, which updated TrueNorth to achieve speeds about 4,000 times faster. However, NorthPole is arguably not a neuromorphic system, and is perhaps better described as very efficient artificial neural network inference engine.
Intel: Loihi and Hala Point
Intel introduced its Loihi neuromorphic system in 2017, claiming a 1,000 x improvement on the neural networks of the time. It used Intel’s 14 nanometre technology and comprised 128 clusters of 1,024 artificial neurons, giving a total of 131,072 simulated neurons, and 130 million synapses – fewer than TrueNorth, and far fewer than the human brain’s 800 trillion synapses.
The upgraded Loihi 2 chip was released in 2021, and in 2024 Intel delivered the first system based on this chip, called Hala Point. The system supports over a billion neurons and 128 billion synapses, consuming 2,600 watts of power. This neuron capacity is claimed to be roughly equivalent to that of an owl brain or the cortex of a capuchin monkey, and the system processes signals 20 times faster than a human brain.
DARPA SyNAPSE
DARPA is the famous research and development agency of the United States Department of Defense, responsible for the development of new technologies for use by the military. It can claim credit for a remarkable range of the technology we use every day, including, at least partly, the internet and the smartphone.
In 2008, DARPA launched SyNAPSE, a program to develop neuromorphic computers with architectures similar to the mammalian brain. SyNAPSE stands for Systems of Neuromorphic Adaptive Plastic Scalable Electronics, and the acronym refers to the junctions between biological neurons. DARPA initially awarded grants of around $5m each to IBM, and to HRL, formerly the research arm of Howards Hughes’ aircraft company. Over the next few years it awarded the companies another $70m or so.
Deep South
In 2024, Andre Van Schaik, director of the International Centre for Neuromorphic Systems at Western Sydney University, claimed that Deep South was the first AI system to match the scale of processing of the human brain. Its 100 FPGA boards can simulate 100 billion neurons generating 220 trillion spikes per second.
An FPGA board is a Field-Programmable Gate Array, which can be programmed like a CPU. But unlike a CPU, when you programme an FPGA you’re changing the layout of the logic gates, which allows them to perform different computations. (Logic gates are the basic elements of a digital computer, performing very simple calculations.) This is a significant difference to SpiNNaker, TrueNorth, Loihi, whose components are mostly fixed.
Deep South’s hardware cost around $2m, and runs on about 40 kilowatts, which is enough to run the air conditioning in a medium-sized building. The most powerful LLMs require the power it takes to run a major city, while the human brain uses 20 watts, enough to power a light bulb.
Deep South is being launched and tested during 2024, and will later be opened up to researchers everywhere, with the design made open source. The aim of the project is to better understand how the brain works by simulating it at scale.
Others
SpiNNaker 2, Loihi 2, and DeepSouth are the main current rivals to succeed the SpiNNaker machine in Manchester as the world’s biggest and fastest neuromorphic computer.
Other significant projects are being pursued at the Institute of Informatics (INI) in Zurich, and the University of Heidelberg. These institutions are creating neuromorphic platforms based primarily upon analogue rather than digital substrates.
Neuromorphic computing conferences
The neuromorphic computing community has two main workshops each year, a three-week event Telluride, Colorado, and a two-week event in Capo Caccia, in Sardinia, Italy.
