Is Neuromorphic Computing Ready for Its Close-Up?

With AI getting daily headlines and computers, phones and smart devices essentially omnipresent, there's a growing awareness that all this computation has a cost — one measured both in dollars and in energy.

Training a large language model (LLM) can cost millions, ranging from $1.4M for a single GPT-3 training session to the $78.4M spent to train GPT-4. The energy impact of training LLMs is even more dramatic: the carbon footprint generated developing GPT-4 is equivalent to the emissions from powering more than 1,300 homes for a year. Sustainable? Not exactly.

Enter Neuromorphic Computing: A Sustainable Alternative

Enter a technology that's been around for decades but is finally finding its way into the spotlight: neuromorphic computing.

Simply put, neuromorphic computing is an approach to hardware and software design that imitates the workings of the human brain, being modeled on the neural and synaptic structures and functions used by the brain to process information. This approach differs radically from the current computing paradigm — and it does so in ways that seem to offer a solution to the sustainability drama now roiling AI.

How Neuromorphic Computing Differs from Traditional AI

Despite the use of the phrase "neural networks" to describe the technologies that have powered recent AI progress, the algorithms and hardware behind current AI systems work very differently from biological neurons.

In human neurology, neurons are messenger cells that relay information within the brain and nervous system through a process known as "spiking" — in which chemical and electrical signals reach a threshold, are released and move through a network of connection points called synapses.

Spiking Neural Networks

This neurological paradigm spawned the concept of spiking neural networks (SNNs) — a type of artificial neural network made up of spiking neurons and synapses that mimic biological neurons and synapses.

In neuromorphic computing systems, individual neurons have their own charge, delay and threshold values; synapses have delay and weight values; and all of these values can be programmed within the system. Among other system features, neuron and synaptic delay values allow for asynchronous dissemination of information.

Crucially, this event-driven spiking paradigm — in which technological neurons echo biological neurons by communicating via spikes of activity rather than the numerical values used by conventional neural networks — is also vastly more energy-efficient than the current approach to computation.

The Advantages of Brain-Inspired Computing

Aiming to make technologies more versatile and adaptable, this approach promises results far beyond traditional architectures.

Unlike traditional computing, which relies on sequential processing, neuromorphic systems use parallel processing for enhanced efficiency and lower power consumption — making them suitable for real-time, adaptive tasks in fields like robotics, healthcare and defense.

Offering ultra-low power, real-time processing for edge devices and high-performance computing, neuromorphic computing is now poised to drive the next wave of AI — mimicking the mind and powering the future.