Explainer
What Is Neuromorphic Computing?
Neuromorphic computing designs chips that borrow ideas from how the brain processes information. The word means brain-shaped. These chips differ from CPUs and GPUs in how they represent data, when they compute, and where they keep memory. They are aimed mainly at tasks that need very low power and fast reactions to sensor data.
The idea behind it
A conventional chip runs on a clock and keeps most of its data in memory that is separate from the processor. Moving data back and forth costs time and energy.
The human brain does something different. Its neurons both store and process information, and they are active only when they have something to signal. The brain uses roughly 20 watts in total, though comparing that with a chip is not straightforward because the tasks differ.
Neuromorphic engineers copy three features of this design: spikes, event-driven operation, and memory placed next to compute.
Spiking neural networks
Most artificial neural networks pass numbers between layers on every step. A spiking neural network, or SNN, works differently. Each neuron collects incoming signals over time, and its internal value, called the membrane potential, rises. It also leaks away slowly. When the value crosses a threshold, the neuron fires a spike and resets.
A spike is a brief event. Information is carried in when spikes occur and how many occur, not only in a continuous value. This matches signals that arrive over time, such as sound, vibration and vision.
Training spiking networks is harder than training standard networks. Spikes are all-or-nothing, which makes the usual gradient method difficult to apply. Researchers use workarounds, such as surrogate gradients, or they train a conventional network and convert it. Tools and software are less mature than those for GPUs.
Event-driven hardware
Event-driven means the chip does work only when an event, such as a spike, arrives. If nothing changes in the input, almost nothing happens and little power is used. Many designs are asynchronous, meaning they have no single global clock.
This saves energy when the data is sparse, which means most of the input is quiet most of the time. If the data is dense and changes constantly, the advantage shrinks. Claims of large energy savings depend heavily on the workload.
Memory is spread across many small cores, each holding its own neurons and connections. This avoids a long trip to a distant memory for every operation.
Chips and systems
IBM's TrueNorth, introduced in 2014, held about one million neurons and 256 million synapses and used roughly 70 milliwatts. A synapse is the connection between two neurons.
Intel's Loihi line began in 2017, and Loihi 2 followed in 2021. In 2024 Intel described Hala Point, a research system built from 1,152 Loihi 2 chips with about 1.15 billion neurons.
SpiNNaker, from the University of Manchester, uses many small Arm cores to simulate spiking networks. Its successor, SpiNNaker2, is commercialised by SpiNNcloud.
BrainChip sells its Akida processor, which is aimed at edge devices. SynSense and Innatera make low-power spiking chips for sensing. Prophesee makes event cameras, whose pixels report only changes in brightness, a natural partner for event-driven processing.
Where neuromorphic chips are used
The strongest cases are always-on sensing at very low power. Examples include detecting a spoken keyword, recognising a gesture, spotting an unusual vibration in a machine, or tracking motion with an event camera.
Other uses are in robotics, where fast reaction times matter, and in research on certain optimisation and sparse-computation problems.
Neuromorphic chips have not replaced GPUs for large language models or image generation. Those workloads are dense and well matched to GPU hardware. Neuromorphic hardware should be judged on its own territory: small, sparse, time-based problems at the edge.
What to watch
Compare a neuromorphic chip with the best conventional low-power chip on the same task, not with a large GPU. The relevant question is whether it meets the accuracy, latency and power target for a real product.
Software support matters as much as silicon. A chip that few engineers can program will struggle to find customers.
Do not confuse neuromorphic computing with biocomputing. Biocomputing uses living neurons grown in a lab as the computing material. Neuromorphic chips are ordinary silicon that imitates some features of the brain.
Companies mentioned
- Intel (Loihi 2) Loihi is Intel's neuromorphic research processor, a chip that imitates the brain by using spiking neural networks, where artificial neurons exchange brief electrical pulses and only consume energy when they fire. This differs from GPUs, which process dense arrays of numbers on every cycle. Loihi 2 was released in September 2021. Intel's Hala Point system combines 1,152 Loihi 2 chips to support up to 1.15 billion neurons, and Intel described it in 2024 as the largest neuromorphic system. Intel is headquartered in Santa Clara, California.
- IBM (TrueNorth/NorthPole) IBM's TrueNorth, unveiled in 2014, is a brain-inspired chip with 4,096 cores that together simulate just over one million neurons and about 268 million synapses, at a power draw IBM put near 70 milliwatts. Neuromorphic designs like this process events as they occur instead of running constant arithmetic. In 2023 IBM described NorthPole, an inference-only chip that places memory beside the compute cores to avoid shuttling data back and forth. IBM is headquartered in Armonk, New York.
- BrainChip BrainChip develops the Akida processor, a neuromorphic design for low-power AI at the edge in devices such as sensors and cameras. Neuromorphic means the chip is modelled loosely on the brain and processes data as sparse events rather than continuously, which can cut energy use compared with conventional processors. The company sells Akida both as licensable IP and as a reference chip, and it supplies a software toolkit called MetaTF for building models. BrainChip is headquartered in Sydney, Australia, and its shares trade on the Australian Securities Exchange under the ticker BRN.
- SynSense SynSense designs neuromorphic chips that run spiking neural networks, a type of model where artificial neurons communicate with brief pulses and consume power mainly when something changes in the input. Its Speck chip pairs a dynamic vision sensor with a processor for always-on camera tasks, and its Xylo chips target audio and other sensor signals. The company was founded in Zurich in 2017, based on research at the Institute of Neuroinformatics of the University of Zurich and ETH Zurich. It also publishes open-source tools named Rockpool and Sinabs for training and deploying such networks.
- Innatera Neuromorphic edge AI startup; T1 microcontroller in mass production with smart-remote design win, Pulsar MCU launched at CES 2026.
- Prophesee Prophesee designs event-based vision sensors. A conventional camera records full frames at fixed intervals, while each pixel in an event-based sensor reports only when its brightness changes, which cuts data volume and delay when scenes move fast. The company supplies the sensors together with its Metavision software tools. Its GenX320 sensor is offered in a kit for the Raspberry Pi, and it has a partnership with Tobii on eye tracking. Prophesee is headquartered in Paris.
- SpiNNcloud Systems SpiNNcloud Systems builds large computers around SpiNNaker2, a chip made of many small processor cores that exchange small messages quickly, an approach used to simulate brain-like networks and run sparse AI workloads. The company is a spin-off from Professor Christian Mayr's research group at TU Dresden in Germany. In April 2025 TU Dresden announced a SpiNNcloud supercomputer based on the chip, comprising 35,000 chips and more than five million processor cores. In 2023 the European Innovation Council awarded the company a 2.5 million euro Transition grant for an edge version of the technology.
- GrAI Matter Labs GrAI Matter Labs developed NeuronFlow, a neuromorphic processor design that does work only on data that has changed, an approach borrowed from how brains handle signals, to run AI at low power on edge devices such as cameras. Its GrAI VIP chip was aimed at vision and audio. Snap Inc. acquired the company in 2023, so it no longer appears to sell chips as an independent business.
Company details come from the Compute Compass directory and may be out of date. This is not investment advice.