By Joseph Moss, International Banker
Although still in early development, neuromorphic computing is attracting considerable interest from technology companies, scientists and researchers as a possible alternative to traditional computing architectures. And with a growing range of potential applications now being identified—from artificial intelligence (AI) and robotics to energy systems and defence—the interest in neuromorphic systems continues to gain traction. That said, significant technical and commercial challenges to its development must first be addressed before determining where and when this technology will ultimately prove most valuable.
Aiming to create hardware that mimics the brain’s neuro-synaptic systems, neuromorphic computing is far from being a new concept. The term was first coined by US scientist and engineer Carver Mead at the California Institute of Technology, who in the 1980s introduced the notion of designing electronic circuits that mimic the analogue behaviour of neurons and synapses, rather than relying solely on binary logic. Although the technology of the time limited practical applications, Mead’s work has been crucial for driving the focus of neuromorphic engineering towards physically implementing brain-like computation in hardware.
The neuromorphic architecture thus stands in stark contrast to the widely used von Neumann framework (introduced by John von Neumann), which distinctly separates memory (where data is stored) and processing (where data is worked on), so that information constantly moves between the two. While this structure has largely proven successful in completing computing tasks to date, its limitations are increasingly being revealed as the complexities of such tasks continue to intensify. This is especially pertinent in the age of AI, with large-scale systems such as data centres already requiring substantial resources to move data.
As the von Neumann architecture’s inefficiencies and shortcomings become more pronounced, the world is gradually turning to neuromorphic computing as an intriguing alternative. Inspired by the human brain, it seeks to more closely integrate memory and processing, specifically by replicating the function of neurons—the brain’s basic units, which both store and process information.
Neurons communicate with each other by being activated (spiked) and releasing electrical signals that travel across a complex network of nodes called synapses. Neuromorphic systems use specialised hardware to model this ecosystem, primarily via spiking neural networks (SNNs), which are artificial networks of spiking neurons and synapses that respond to various inputs.
Rather than processing instructions sequentially, therefore, these systems can process information simultaneously across many nodes. This distinction means that neuromorphic systems may require less data to learn effectively, be better positioned to adjust to changing environments and perform certain tasks with significantly less power.
“The idea of neuromorphic computing is to take inspiration of the brain for designing computer chips that merge memory and processing. In the brain, synapses provide a direct memory access to the neurons that process information,” Julie Grollier, a specialist in neuromorphic physics, explained to the science journal Nature in October 2019. “That’s how the brain achieves impressive computational power and speed with very little power consumption. By imitating this architecture, neuromorphic computing provides a path to building smart chips that consume very little energy and, meanwhile, compute fast.”
Indeed, it is perhaps the concerns about the copious energy consumption of the AI revolution—which still mostly runs on conventional hardware and uses large datasets and complex mathematical models to identify patterns and make predictions—that is prompting greater support for neuromorphic computing’s development. Training and running large AI models require progressively greater amounts of energy, particularly in data centres, which are now consistently home to the most computationally demanding tasks of any technological field.
With neuromorphic systems processing information in a more distributed manner and consuming energy only when activity occurs, they could register potentially substantial improvements in energy efficiency for certain tasks. “While there still isn’t a killer app…there are lots of areas where neuromorphic computing will provide significant gains in energy efficiency and performance, and I’m sure we’ll start to see wide adoption of the technology as it matures,” Tony Kenyon, a professor of nanoelectronic and nanophotonic materials at University College London, told the British Broadcasting Corporation (BBC) in June 2024.
Another factor driving interest is the growth of edge computing, in which data is processed near where it is generated rather than being sent to centralised servers. Applications such as autonomous vehicles, industrial sensors and wearable devices require fast, local decision-making with minimal energy consumption. Traditional computing architectures are not always well suited to these constraints.
With that in mind, neuromorphic systems are increasingly regarded as being more attuned to the needs of edge computing. Their ability to process information quickly and efficiently without relying on constant data transfer makes them attractive for decentralised applications. “The characteristics of neuromorphic architecture make it suitable for edge AI,” according to IBM. “Its low power consumption can help with the short battery life of devices like smartphones and wearables, while its adaptability and event-driven nature fit the information processing methods of remote sensors, drones and other Internet of Things (IoT) devices.”
Autonomous systems may also benefit from neuromorphic architecture, especially given that the likes of self-driving vehicles and robotics require rapid processing of sensory information to enable faster reaction times and more efficient real-time decision-making across dynamic environments. “Because of its high performance and orders of magnitude gains in energy efficiency, neuromorphic computing can help improve an autonomous vehicle’s navigational skills, allowing for quicker course correction and improved collision avoidance while lowering energy emissions,” IBM also observed.
Among neuromorphic hardware in development, Intel’s Loihi chip represents one of the most prominent examples. It has been designed to emulate the behaviour of biological neural networks, employing SNNs to transmit information through discrete electrical pulses, thereby mimicking how neurons communicate in the brain. This allows the system to process information in a highly event-driven manner, consuming energy only when activity occurs.
What makes Loihi particularly momentous, however, is the concrete evidence that a major technology company is willing to invest in a fundamentally different and somewhat unproven computing paradigm. What’s more, the second-generation version, Loihi 2, is now being deployed in research environments and tested across a range of applications, including robotics, optimisation problems and adaptive control systems. According to Intel, Loihi 2 “outperforms its predecessor by up to 10x and comes with an open-source, community-driven neuromorphic computing framework called Lava”.
Early results suggest that for certain types of tasks, neuromorphic hardware such as Loihi can achieve substantial efficiency gains compared with traditional processors.
Early results suggest that for certain types of tasks, neuromorphic hardware such as Loihi can achieve substantial efficiency gains compared with traditional processors. However, these advantages remain context-specific, and Loihi is still largely confined to experimental and research settings. As such, its development underscores both the abundant promise of neuromorphic computing and the distance yet to be covered before widespread commercial deployment becomes viable.
Indeed, despite that promise, neuromorphic computing faces several significant challenges. Unlike traditional computing, which has benefited from decades of software development, neuromorphic systems lack a mature ecosystem. Programming these systems requires new approaches and tools, which are still under exploration. At the same time, the realms of traditional computing and AI continue to improve markedly. As such, advances in specialised AI hardware, such as GPUs (graphics processing units) and AI accelerators, may address some of the same challenges that neuromorphic computing aims to solve.
With most neuromorphic technologies still in experimental or early-deployment phases, substantial investments will be required not only to scale their production but also to integrate these systems into existing infrastructure. And while the potential applications of neuromorphic computing appear broad, it remains unclear at this stage which use cases will prove commercially viable at scale. In turn, companies may remain hesitant to commit large resources to this space for the time being.
Such challenges to neuromorphic computing mean that it remains an emerging technology—for now, at least. But as the constraints of traditional computing become more apparent and problematic, its relevance and appeal should begin to grow more meaningfully. By striving to rethink how machines process information, its real promise lies in delivering more efficient, adaptable and responsive computing systems.
Translating biological principles into reliable, scalable technology in a sustainable way appears to be the next step. Carver Mead believes that being bold is vital for achieving progress. “Along with learning how the brain does what it does, we also need people to be taking some risks” in terms of advanced neuromorphic computer architectures, he told Photonics Focus magazine in September 2024. “You can’t learn without doing. You can’t learn how a thing works unless you can build it and make it work. And that is a tall order.”