Neuromorphic computing explained how brain inspired chips work

[CarlosBuddle]

TL;DR

• Neuromorphic computing builds chips that mimic aspects of the brain, using spiking neurons, local memory and event driven processing instead of a central clock and separate memory
• The goal is very low power computing for tasks such as sensing, pattern recognition and real time control, especially at the edge
• Intel's Hala Point system, unveiled in April 2024, packs 1,152 Loihi 2 processors with about 1.15 billion artificial neurons, and IBM's NorthPole chip was described in Science in October 2023
• Sandia National Laboratories reported in 2026 that a fruit fly brain simulation on 12 Loihi 2 chips ran more than 100 times faster than conventional numerical simulation
• Neuromorphic hardware is promising but still niche: software tools, programming models and proven commercial applications are the main hurdles

Why look to the brain

The human brain runs on roughly 20 watts, about the power of a dim light bulb, yet it recognises faces, understands language and coordinates movement in real time. Modern AI systems that perform some of these tasks can consume vastly more energy, especially when training and serving large models in data centres. That gap has inspired researchers for decades. If computers worked more like brains, could they be far more efficient

Neuromorphic computing tries to answer that question. The term was coined by Caltech engineer Carver Mead in the late 1980s to describe electronic circuits that mimic neural systems. Today it covers a range of chips and systems that borrow ideas from neuroscience, from spiking neurons to massively parallel, event driven communication. The field has grown steadily, with major contributions from universities, national laboratories and companies

This thread explains how neuromorphic chips differ from conventional processors, surveys leading systems and research results, and discusses where the technology might fit. It also offers a balanced view of the challenges. Neuromorphic computing is exciting, but it is not about to replace GPUs for mainstream AI. Understanding why helps set realistic expectations

How conventional computers work

Most computers follow the von Neumann architecture, in which a processor fetches data and instructions from separate memory, processes them and writes results back. A central clock ticks billions of times per second, and every part of the chip does work on every tick, whether or not there is anything useful to compute. Moving data between memory and processor consumes much of the energy in modern systems. This bottleneck matters especially for AI, which shuffles huge amounts of data

GPUs and AI accelerators improve on this by performing many operations in parallel and placing memory closer to compute. They are extraordinarily effective for the dense matrix arithmetic at the heart of today's neural networks. However, they still move large volumes of data and run continuously while active. Energy use scales with the size of models and the volume of requests

How neuromorphic chips differ

Neuromorphic chips typically use spiking neural networks. Instead of passing continuous numbers between neurons every cycle, artificial neurons accumulate input and fire a brief spike only when a threshold is reached. If nothing changes, little or nothing happens. This event driven behaviour means energy is spent mainly when there is information to process

Memory and computation are co located. Each artificial neuron and synapse stores its state locally, so data does not need to travel back and forth to a distant memory bank. This mirrors the way biological synapses both store and process information. It avoids much of the von Neumann bottleneck

Communication is massively parallel and asynchronous. Neurons send spikes to many others across a network on chip, often without a global clock coordinating every step. Chips can be tiled together to build larger systems. The result is hardware well suited to sparse, time varying signals such as those from sensors

Learning can happen on chip in some designs. Biologically inspired learning rules let synapse strengths adjust based on the timing of spikes, allowing systems to adapt in real time. This capability is attractive for robots and devices that must learn from their environment without connecting to the cloud. It remains an active research area

Leading systems

Intel has been one of the most active companies in the field. Its Loihi 2 research chip, introduced in 2021, supports programmable spiking neurons and on chip learning. In April 2024, Intel unveiled Hala Point, a system built from 1,152 Loihi 2 processors with about 1.15 billion neurons and 128 billion synapses, deployed at Sandia National Laboratories. Intel described it as the world's largest neuromorphic system at the time

IBM took a related but distinct approach with NorthPole, described in the journal Science in October 2023. NorthPole is not a spiking chip but is inspired by the brain's integration of memory and computation, placing memory directly alongside compute cores. IBM reported strong energy efficiency and speed for image recognition inference compared with conventional chips. It shows how brain inspired principles can apply beyond spiking networks

In Europe, the SpiNNaker project, led by the University of Manchester, built a massively parallel computer from many simple processor cores designed to simulate spiking neurons. Its successor, SpiNNaker2, developed with Technische Universitat Dresden and the spin out SpiNNcloud Systems, scales the approach further. Sandia also hosts a SpiNNaker2 based system, giving US researchers access to two different neuromorphic architectures. The Human Brain Project in the EU supported much of this European work

Startups are targeting edge applications. Companies such as BrainChip, with its Akida processors, and Innatera, with ultra low power spiking processors for sensors, aim to bring neuromorphic computing into devices such as wearables, cameras and industrial sensors. Their pitch centres on always on processing at very low power. Commercial traction is growing but still early

Recent research results

National laboratories are exploring neuromorphic computing for scientific workloads. In February 2026, Sandia National Laboratories described work on its Hala Point and SpiNNaker2 systems, reporting that a simulation of the fruit fly connectome running on 12 Loihi 2 chips ran more than 100 times faster than numerical simulation on conventional hardware, with the biggest advantages for networks with sparse activity. Sandia researchers have also explored using neuromorphic hardware for random walk methods relevant to physics problems such as heat flow. These results suggest neuromorphic systems may find niches in scientific computing as well as AI

Research on spiking neural networks for AI continues too. Techniques for converting conventional neural networks into spiking versions, and for training spiking networks directly, have improved. Accuracy on standard benchmarks has narrowed the gap with conventional networks for some tasks. Efficiency gains are most convincing when inputs are naturally sparse or event based

Event based sensors are a natural partner. Neuromorphic cameras, also called event cameras, report only changes in brightness at each pixel rather than full frames, producing sparse streams of events. Pairing them with neuromorphic processors can enable fast, low power vision for robotics, drones and automotive systems. This combination is one of the most promising application areas

Where neuromorphic computing might fit

Edge devices are the most likely early winners. Wearables, hearing aids, smart sensors and battery powered cameras need to process signals continuously with minimal energy. Neuromorphic processors can listen for keywords, detect anomalies or recognise gestures while consuming tiny amounts of power. That could extend battery life and reduce the need to send data to the cloud

Robotics and autonomous systems are another fit. Robots must react quickly to changing environments, and neuromorphic systems' low latency and on chip learning suit such tasks. Research groups have demonstrated neuromorphic control of robotic arms and drones. Real world deployment will depend on robustness and integration with existing control software

Scientific computing and brain simulation round out the picture. Neuroscientists use neuromorphic systems to simulate large networks of neurons, testing theories about brain function. National laboratories explore them for specific algorithms. These uses may remain specialised but are scientifically valuable

The challenges

Software is the biggest hurdle. Programming spiking neural networks is very different from programming conventional neural networks, and tools are less mature. Intel's open source Lava framework and other projects aim to make development easier, but the ecosystem is small compared with mainstream AI frameworks. Without good tools, developers stick with GPUs

Benchmarks and comparisons are tricky. Neuromorphic systems often shine on specific tasks under specific conditions, and fair comparisons with optimised conventional hardware require care. Some reported efficiency gains shrink when compared against the latest GPUs or specialised accelerators. Clear, independent benchmarking would help the field

Commercial scale remains limited. Large research systems exist, but volume products are mostly in early stages. Manufacturing, supply chains and customer adoption take time. The field may follow the path of other technologies that spent years in niches before finding broad applications

How it relates to quantum computing

Neuromorphic and quantum computing are sometimes mentioned together as beyond conventional computing approaches, but they are very different. Neuromorphic chips are classical devices built from ordinary transistors, designed for efficiency on certain tasks. Quantum computers exploit quantum mechanics for fundamentally different kinds of computation. Neuromorphic computing poses no threat to cryptography

Both fields share a challenge, however. Each must show clear advantages over rapidly improving conventional hardware. GPUs and AI accelerators keep getting faster and more efficient, raising the bar. Niche advantages must be compelling enough to justify new tools and ecosystems

Closing thoughts

Neuromorphic computing offers an intriguing vision of hardware that thinks more like a brain, sipping power while reacting in real time. Large systems such as Hala Point and SpiNNaker2, chips such as NorthPole and edge processors from startups show steady progress. Recent results from Sandia suggest promising niches in science as well as AI. The main barriers are software, benchmarks and commercial scale, and they are being worked on actively

If you work with neuromorphic hardware or event based sensors, please share your experiences below. Questions from newcomers are very welcome too. This is a field where curiosity and cross disciplinary thinking pay off


Key references for citation

• Mead, Neuromorphic electronic systems, Proceedings of the IEEE, 1990
• Intel, Hala Point neuromorphic system announcement, April 2024
• Modha and colleagues, Neural inference at the frontier of energy, space, and time (NorthPole), Science, October 2023
• Sandia National Laboratories, neuromorphic computing research update, February 2026
• SpiNNaker and SpiNNaker2 projects, University of Manchester and TU Dresden