The AI Energy Crisis and a Biological Solution

Artificial intelligence is transforming industries, from healthcare and finance to manufacturing and transport. However, this progress comes with a growing challenge: the energy required to train and run modern AI systems. Large language models (LLMs), cloud-based services and increasingly complex neural networks place substantial demands on data centres, electricity supplies and cooling infrastructure.

As AI becomes more widely deployed, improving AI energy efficiency is no longer simply a technical objective. It is an important part of building sustainable digital infrastructure.

One promising alternative is neuromorphic computing, an approach to hardware design inspired by the way biological brains process information. Instead of relying exclusively on conventional processors and continuously executed neural network operations, neuromorphic systems use specialised circuits designed to process signals efficiently.

Some neuromorphic benchmarks report substantial energy savings, potentially reaching 90% or more for particular workloads compared with conventional computing approaches. However, these results depend on the task, hardware, software and measurement method. They do not mean that every AI model automatically uses 90% less energy.

The Bottleneck of Traditional AI Hardware

The Von Neumann Bottleneck

Most conventional computers use an architecture in which processing units and memory are physically separated. CPUs and GPUs repeatedly move data between memory and processing circuits, consuming energy and time.

This problem is often called the Von Neumann bottleneck. In AI workloads, transferring model parameters, intermediate results and input data can account for a significant proportion of total energy consumption.

GPUs remain exceptionally effective at performing parallel mathematical operations. Nevertheless, their architecture is not always ideal for applications that require continuous, low-latency responses to sparse or rapidly changing sensor signals.

Continuous Computation and Energy Waste

Many conventional artificial neural networks (ANNs) process inputs through dense numerical operations. Depending on the model and implementation, these computations may continue even when relatively little new information is available.

For applications such as event-based vision, wearable sensors and industrial monitoring, this can create unnecessary processing overhead. A device may need to detect only a sudden movement, an unusual vibration or a change in heart rhythm, rather than repeatedly analysing an unchanged stream.

The Sustainability Challenge

Scaling conventional AI infrastructure increases electricity demand, operating costs and cooling requirements. Although more efficient GPUs, quantisation and optimised software can reduce energy consumption, further improvements are needed, particularly for battery-powered and always-on devices.

Neuromorphic computing offers a different approach by redesigning how information is represented, processed and stored.

What Is Neuromorphic Computing?

Mimicking the Human Brain

Neuromorphic computing takes inspiration from biological neurons and synapses. Rather than simply reproducing the brain's biology, engineers use these principles to design electronic systems that process information through networks of artificial neurons.

Some architectures combine computation and memory within the same physical structures or place them close together. This can reduce data movement, although the exact implementation varies between chips.

Examples include Intel's Loihi family, IBM's TrueNorth research chip and BrainChip's Akida platform. Each uses its own design and supported computational methods, so their performance should not be treated as directly interchangeable.

Spiking Neural Networks Explained

A central concept in neuromorphic computing is the spiking neural network (SNN). Unlike conventional neural networks that typically represent activations as numerical values, SNNs communicate through discrete events known as spikes.

An artificial neuron accumulates input signals and emits a spike when certain conditions are met. Information can therefore be represented through the timing and pattern of these events.

In suitable applications, many neurons remain inactive when there is no relevant input. This event-driven behaviour can reduce unnecessary computation and energy use.

However, not every SNN is automatically energy-efficient. Network activity, neuron design, memory access, communication overhead and the hardware implementation all influence actual power consumption.

Bringing Memory and Computation Closer Together

Neuromorphic chips often distribute memory and processing across many computational units. This reduces the need to move every intermediate value through a central memory system.

Combined with event-driven processing, this design can improve efficiency for workloads involving sparse signals, temporal patterns and continuous sensor input.

Breaking Down the Benchmarks: The 90% Efficiency Leap

The headline claim of 90% lower energy consumption needs careful interpretation. Neuromorphic computing research has demonstrated impressive efficiency on selected tasks, but there is no single universal benchmark proving that all neuromorphic chips outperform all GPUs by this margin.

A fair comparison measures the same task, with comparable accuracy, equivalent input conditions and consistent accounting for the entire system.

Neuromorphic Hardware Comparison

Understanding Energy per Inference

Energy per inference measures how much energy a system consumes to produce one model output. It is commonly expressed in joules per inference or millijoules per inference.

For a simplified measurement, energy can be calculated as:

Energy = Average power × Processing time

For example, a hypothetical system operating at 2 watts for 10 milliseconds consumes 0.02 joules, or 20 millijoules, during that interval.

This calculation illustrates why both power and latency matter. A low-power chip may still consume considerable energy if it takes too long to complete a task. Conversely, a powerful GPU can finish a batch quickly, potentially making its energy per inference competitive under suitable conditions.

Meaningful neuromorphic hardware benchmarks should also report model accuracy, throughput, batch size, idle power and whether host processors or external memory are included in the measurement.

Latency and the 90% Metric

Neuromorphic systems can deliver very low latency for particular event-driven workloads, sometimes operating within a milliwatt-scale power budget. This is particularly attractive when devices must react immediately to changing sensor inputs.

A 90% energy reduction means that the neuromorphic implementation consumes one-tenth of the energy used by the chosen baseline. Such a result may be achievable in favourable scenarios, but it should be supported by a published benchmark with a clearly defined comparison.

The advantage is generally most plausible when inputs are sparse, activity is intermittent and conventional processing would otherwise perform unnecessary work. Dense batch processing, large language model inference and other highly parallel workloads may favour conventional GPUs or require a different architectural approach.

The key takeaway is that energy savings must be measured for a specific workload rather than assumed from the chip's design alone.

Advantages Beyond Extreme Energy Efficiency

Real-Time Processing

Neuromorphic systems can process incoming events as they arrive, making them suitable for applications requiring immediate responses. Examples include detecting collisions, identifying sudden movements and recognising unusual machine vibrations.

On-Device Learning

Some neuromorphic architectures support local adaptation or learning mechanisms. This creates opportunities for devices to respond to changing environments without continuously sending raw data to cloud servers.

Nevertheless, on-device learning is not a universal feature of every neuromorphic chip. Some platforms primarily target inference, while training and model development may still take place on conventional hardware.

Robustness to Noise and Component Failures

Brain-inspired systems can be designed to tolerate noisy signals and distributed computation. Their parallel structure may offer useful resilience in certain applications, particularly when information is encoded across multiple neurons.

However, fault tolerance is not guaranteed simply because a system uses SNNs. It depends on the network design, training process, hardware reliability and application requirements.

Real-World Applications for Brain-Inspired Chips

Autonomous Vehicles and Drones

Event-based cameras capture changes in a scene rather than producing only conventional full-frame images. Neuromorphic processors can analyse these signals efficiently, potentially helping drones and autonomous machines detect movement with low latency.

Safety-critical driving decisions still require rigorous testing, redundancy and reliable integration with other perception systems.

Smart Healthcare Wearables

Wearable devices can monitor heart rhythms, movement and other physiological signals continuously. Neuromorphic processing may help identify unusual patterns while reducing energy consumption and extending battery life.

Clinical applications must also meet requirements for accuracy, privacy, validation and regulatory compliance.

Industrial IoT and Robotics

Factories generate streams of vibration, pressure, sound and visual data. Neuromorphic systems could detect emerging equipment faults or unexpected movements without requiring every sensor reading to be processed by a large central computer.

This can reduce communication overhead and potentially lower operating costs in distributed industrial environments.

Challenges and the Road to Mainstream Adoption

Despite their promise, neuromorphic systems face significant barriers.

Software ecosystem: Conventional frameworks such as PyTorch and TensorFlow support extensive development workflows. SNNs often require specialised simulators, conversion tools and hardware-specific programming environments, making development less straightforward.

Model conversion: Converting a conventional ANN into an SNN may reduce compatibility or accuracy. Researchers are improving training methods, quantisation and conversion techniques, but results vary between tasks.

Manufacturing and deployment: Specialised processors require investment in hardware design, production, developer tools and system integration. Organisations must evaluate these costs alongside potential electricity savings.

Benchmark consistency: Comparing a neuromorphic chip with a GPU is difficult when they use different model architectures or achieve different accuracy levels. Independent, reproducible benchmarks are essential for trustworthy claims.

Hybrid systems may provide a practical route forward, using GPUs or CPUs for complex workloads and neuromorphic processors for sparse, low-latency sensor processing.

Paving the Way for Sustainable Artificial Intelligence

Neuromorphic computing offers an important direction for sustainable AI. By combining spiking neural networks, event-driven computation and memory-efficient hardware, it can reduce energy consumption for selected applications while enabling fast, continuous processing.

The reported potential for 90% energy savings is promising, but it should be understood as a workload-dependent result rather than a universal guarantee. The strongest opportunities currently lie in edge AI, event-based vision, sensor processing and other applications where low power and immediate responses are more important than running the largest possible model.

Over time, improved software tools, better training methods and more accessible hardware could expand adoption across robotics, healthcare, industrial monitoring and autonomous systems. Neuromorphic processors are unlikely to replace conventional GPUs everywhere, but they could become an important part of a more diverse and energy-efficient AI infrastructure.

Which industry do you think will benefit most from neuromorphic computing: healthcare, robotics, autonomous vehicles or industrial IoT? Share your thoughts in the comments and subscribe to our newsletter for more insights into sustainable technology, AI hardware and emerging computing architectures.