What is neuromorphic computing? Brain-inspired chips and how they compare with living neurons

Neuromorphic computing is an approach to building computer hardware that mimics the structure and dynamics of biological nervous systems. Neuromorphic chips typically use spiking artificial neurons, event-driven communication and memory placed next to computation. The goal is to process sensory, time-varying data with much lower energy and latency than conventional processors, which shuttle data between separate memory and logic.

Key takeaways

• Neuromorphic computing emulates neurons and synapses in electronics, usually using spikes rather than continuous values.
• Carver Mead’s 1990 paper in Proceedings of the IEEE helped define the field.
• Notable chips include IBM’s TrueNorth (2014), Intel’s Loihi (2018) and the SpiNNaker system from the University of Manchester.
• Neuromorphic systems are programmed with spiking neural networks and often support local learning rules.
• It is a silicon cousin of organoid intelligence, which computes with real neurons instead of imitations.

How does neuromorphic computing work? #

Conventional computers follow the von Neumann design: a processor executes instructions on a clock and fetches data from separate memory. Moving data back and forth costs time and energy, a constraint often called the von Neumann bottleneck. Neuromorphic designs depart from this in several ways.

1. Spiking neurons. Each artificial neuron integrates incoming signals and emits a discrete spike when a threshold is crossed, like a biological action potential.
2. Event-driven operation. Circuits do work only when spikes arrive. If nothing changes, little power is used.
3. Co-located memory and compute. Synaptic weights are stored right next to the neurons that use them.
4. Massive parallelism. Many simple cores or circuits operate simultaneously rather than one fast processor doing everything in sequence.
5. Local learning. Some chips can update synapses on-chip using rules like spike-timing-dependent plasticity.

Analog vs digital neuromorphic designs #

Mead’s original vision used analog transistor circuits operating in a regime where their physics resembles ion channel behavior. Indiveri et al. (2011, Frontiers in Neuroscience) reviewed many such silicon neuron circuits. Analog designs can be very efficient but suffer from device mismatch and noise. Digital designs like TrueNorth and Loihi are more predictable and easier to manufacture at scale, while still keeping the event-driven, spiking style.

What are the main neuromorphic chips? #

Merolla et al. reported in Science (2014) that TrueNorth achieved its scale with very low power consumption compared with conventional hardware running similar networks. Roy, Jaiswal and Panda (2019, Nature) and Schuman et al. (2022, Nature Computational Science) reviewed the algorithms and applications that suit such hardware, including sensory processing, robotics and optimization.

What are neuromorphic chips used for? #

Because they are best at sparse, time-varying signals, neuromorphic systems are usually explored for workloads where data arrive as streams and decisions must be made quickly with little power. Reviews by Roy et al. (2019) and Schuman et al. (2022) discuss applications such as:

• Sensory processing, including pairing with event-based cameras that report only changes in brightness.
• Robotics and control, where low latency and low power matter more than peak accuracy.
• Optimization and graph problems, which some researchers map onto networks of spiking units.
• Neuroscience simulation, which was a founding goal of SpiNNaker: running large spiking models in biological real time.

In each case the question is the same one biocomputing faces: does a brain-like substrate solve a real problem better than a conventional processor, once energy, accuracy and engineering effort are all counted?

Neuromorphic computing vs organoid intelligence vs conventional AI #

The comparison shows why the fields complement each other. Neuromorphic engineering offers stable, programmable hardware. Biocomputing and wetware computing offer genuine biological complexity, plasticity and potential energy efficiency, which we discuss in biocomputing vs silicon: why living neurons could fix AI’s energy problem.

Why does neuromorphic computing matter for biocomputing? #

There are several direct connections.

• Shared language. Both fields think in spikes, timing and local plasticity. Algorithms developed for neuromorphic chips can inform how we encode and decode signals in living networks.
• Shared frameworks. Reservoir computing, which needs no training of the core system, has been explored on neuromorphic hardware and on brain organoids.
• Hybrid systems. A neuromorphic processor could sit between a microelectrode array and a culture, handling low-latency spike detection and stimulation decisions.
• Benchmarks. Neuromorphic chips provide a useful baseline: if an organoid system is claimed to be efficient, it should be compared with the best brain-inspired silicon, not only with GPUs.

What are the limitations of neuromorphic computing? #

• Software ecosystem. Programming tools and trained models lag far behind deep learning frameworks.
• Training difficulty. Spiking networks are harder to train than conventional networks, although methods like surrogate gradients have improved this.
• Task fit. Gains are clearest on sparse, event-driven workloads, and less clear on dense numerical computation.
• Approximation of biology. Neuromorphic neurons are greatly simplified compared with real neurons and synapses.

These limits are part of why some researchers are exploring living neurons directly. Our article on the future of biocomputing looks at where the approaches may converge, and at Organoid Programming we are building software that treats living tissue as a programmable substrate.

Conclusion #

Neuromorphic computing builds brain-inspired chips that compute with spikes, events and local memory. From Mead’s analog circuits to TrueNorth, Loihi and SpiNNaker, it has shown that borrowing the brain’s architecture can cut energy use for suitable tasks. Neuromorphic computing and organoid intelligence approach the same question from opposite sides: imitating neurons versus using them.

Frequently asked questions

What is neuromorphic computing in simple terms?

Neuromorphic computing means building computer chips that work more like a brain. Instead of a central processor fetching data from separate memory on a fixed clock, neuromorphic chips use many simple artificial neurons that send spikes to each other only when something happens. The aim is to process sensory and time-varying information efficiently, using much less power than conventional chips for suitable tasks.

Who invented neuromorphic computing?

The field is closely associated with Carver Mead at Caltech, who proposed using analog transistor circuits to emulate neural processing. His 1990 paper Neuromorphic Electronic Systems in Proceedings of the IEEE helped define the field and its motivation: that biological systems compute far more efficiently than digital machines on many perceptual tasks, and that electronics could learn from them.

What are examples of neuromorphic chips?

Well-known examples include IBM’s TrueNorth, described in Science in 2014 as a million spiking-neuron chip; Intel’s Loihi, described in IEEE Micro in 2018 with support for on-chip learning; and the University of Manchester’s SpiNNaker system, which uses many ARM processor cores to simulate large spiking networks in real time. Many research groups also build analog and mixed-signal neuromorphic circuits.

How is neuromorphic computing different from organoid intelligence?

Neuromorphic computing imitates neurons in silicon. Organoid intelligence uses real human neurons grown as brain organoids. Neuromorphic chips are manufacturable, stable and programmable but only approximate biology. Organoids have genuine biological learning and complexity but are variable, need life support and are harder to read and control. The two fields share ideas such as spiking codes and local plasticity.

Is neuromorphic computing more energy efficient than GPUs?

For some workloads, particularly sparse, event-driven tasks such as processing sensor streams, neuromorphic chips have been reported to use much less energy than conventional processors. For dense workloads like training large deep learning models, GPUs remain the standard. Efficiency depends strongly on the task, the algorithm and how the comparison is measured.

Sources

1. Mead (1990). Neuromorphic electronic systems. Proceedings of the IEEE.
2. Merolla et al. (2014). A million spiking-neuron integrated circuit with a scalable communication network and interface. Science.
3. Davies et al. (2018). Loihi: a neuromorphic manycore processor with on-chip learning. IEEE Micro.
4. Furber et al. (2014). The SpiNNaker project. Proceedings of the IEEE.
5. Roy, Jaiswal & Panda (2019). Towards spike-based machine intelligence with neuromorphic computing. Nature.
6. Schuman et al. (2022). Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science.
7. Indiveri et al. (2011). Neuromorphic silicon neuron circuits. Frontiers in Neuroscience.