Neuromorphic Computing

Neuromorphic computing is an approach to designing computer hardware and algorithms that imitate the architecture and operating principles of biological brains. Instead of the conventional model in which processing and memory are separated and computation proceeds in synchronised clock cycles, neuromorphic systems co-locate memory and computation and process information through discrete electrical events called spikes. The aim is to achieve large gains in energy efficiency and to enable new forms of low-power, real-time intelligence, particularly for tasks involving sensory data and continuous learning.

Spiking neural networks

The computational model underlying most neuromorphic hardware is the spiking neural network. Unlike conventional artificial neural networks, whose neurons output continuous numerical values at every layer, spiking neurons communicate through brief, discrete pulses that occur only when a neuron's internal state crosses a threshold. Information is encoded in the timing and frequency of these spikes, mirroring how biological neurons signal one another.

This event-driven style of computation is inherently sparse: a neuron consumes energy only when it spikes, rather than continuously. For workloads with naturally sparse or temporally structured data, such as audio, vision from event cameras, and sensor streams, this can translate into dramatic reductions in power consumption compared with running equivalent models on conventional processors.

Hardware

Several research chips embody the neuromorphic approach. Intel's Loihi, introduced in 2017, integrated many neuromorphic cores supporting on-chip learning, with the original design modelling on the order of one hundred thousand neurons and over one hundred million synapses. Its successor, Loihi 2, offers substantially faster processing, far greater neuron density of up to one million neurons per chip, and improved energy efficiency, and supports a wider range of neuron models and learning rules. IBM's TrueNorth and a number of academic platforms, including SpiNNaker and BrainScaleS, pursue related goals with different design choices.

| Feature | Conventional (von Neumann) | Neuromorphic | | --- | --- | --- | | Memory and compute | Separated | Co-located | | Timing | Clock-synchronised | Event-driven | | Signalling | Continuous values | Discrete spikes | | Energy profile | Constant when active | Sparse, spike-triggered |

Applications and challenges

Neuromorphic computing is most promising for edge applications where power is limited and latency matters, including always-on sensing, robotics, gesture and keyword recognition, and adaptive control. Its efficiency aligns it closely with edge AI and TinyML, where models must run on small, battery-powered devices. The field faces significant challenges, however. Programming spiking systems requires different tools and training methods than mainstream deep learning, mature software ecosystems are still emerging, and many results remain at the research stage rather than in volume production. Whether neuromorphic hardware becomes mainstream or remains a specialised complement to conventional accelerators is an open question.

Neuromorphic computing intersects with two of Malaysia's strategic strengths: semiconductor manufacturing and the push toward energy-efficient edge AI. Malaysia is a major node in the global semiconductor supply chain, with extensive assembly, packaging, and testing operations in Penang and Kulim, and the National Semiconductor Strategy aims to move the country up the value chain toward higher-value design and advanced packaging. Specialised chips such as neuromorphic processors depend on exactly these advanced packaging and integration capabilities, giving Malaysian facilities a potential role in their production.

Research interest exists in the academic sector. Universities including Universiti Sains Malaysia, Universiti Teknologi Malaysia, and Universiti Malaya conduct work in microelectronics, embedded systems, and neural computation, and MIMOS, the national applied research institute, works on microelectronics and IC design relevant to next-generation hardware. Collaboration with multinational semiconductor firms operating in Malaysia, such as Intel, which has a long-established presence in Penang, connects local capability to global neuromorphic research.

The practical pull for energy-efficient AI is strong. As Malaysian industries deploy sensors and edge devices in manufacturing, agriculture, and smart city projects under the MyDigital blueprint, low-power on-device intelligence becomes attractive for reducing reliance on connectivity and cloud compute. This aligns with sustainability goals and the broader interest in green AI given the energy demands of growing data centre capacity in Johor and Cyberjaya.

For Malaysia, the nearer-term opportunity lies less in inventing neuromorphic architectures and more in manufacturing the advanced chips that embody them and in applying energy-efficient edge AI across its industrial base.

1. Davies, M., et al. (2018). Loihi: A Neuromorphic Manycore Processor with On-Chip Learning. IEEE Micro.
2. Intel Labs. Taking Neuromorphic Computing to the Next Level with Loihi 2.
3. Maass, W. (1997). Networks of Spiking Neurons: The Third Generation of Neural Network Models. Neural Networks.
4. Open Neuromorphic. A Look at Loihi 2 — Intel Neuromorphic Chip.