The concept sounds futuristic, but the research is decades old and the hardware is increasingly real. Understanding what these chips actually do – and why they matter – gives you a clearer picture of where AI hardware is heading next.
The Problem With How AI Hardware Works Today
To understand neuromorphic chips, it helps to understand what they're reacting against. Most AI today runs on GPUs (graphics processing units) or specialized chips like Google's TPUs. These are very good at one thing: performing billions of matrix multiplications in parallel. That's the mathematical operation at the heart of most modern neural networks, and these chips do it fast.
The problem is the energy cost. Conventional processors are built on what's called the von Neumann architecture – a design where memory and processing are physically separate, which means data has to move back and forth constantly between the two. That movement consumes significant power. When you scale up to AI workloads, which require enormous amounts of data to move through enormous numbers of calculations, the energy bill climbs fast. This is sometimes called the "memory wall" – the bottleneck between compute and memory that limits both speed and efficiency at scale.
GPUs also work in a highly synchronous way: they process data in discrete steps, ticking through operations in lockstep with a clock signal. That structure is efficient for certain tasks but not for the kind of sparse, event-driven processing that biological brains do naturally.
What "Neuromorphic" Actually Means
The word comes from the Greek neuron (nerve) and morphe (form). Neuromorphic computing is, at its core, an attempt to build hardware that works more like a brain than like a conventional processor.
The key idea is spiking neural networks (SNNs). In a biological brain, neurons don't fire constantly – they produce brief electrical spikes only when there's something to respond to, and stay quiet the rest of the time. This is fundamentally different from how conventional artificial neural networks work, where every node in the network is active during every computation, regardless of whether it has anything meaningful to contribute.
Neuromorphic chips are designed to run spiking neural networks natively. They process information in an event-driven way – neurons on the chip "fire" only when their input crosses a certain threshold, just like biological neurons. Because most neurons are inactive most of the time, the chip uses only a fraction of the power it would if everything were running continuously. The memory and processing are also co-located on the chip rather than separated, which eliminates the constant back-and-forth that makes conventional architectures so power-hungry.
The result, in principle, is a chip that can perform certain types of computation at a fraction of the energy cost of a conventional processor.
The Chips That Actually Exist
This isn't purely theoretical. Several significant neuromorphic chips have been built and studied over the past decade.
Intel's Loihi chip, first released in 2018 and followed by the more capable Loihi 2 in 2021, is one of the most well-documented examples. Loihi 2 contains around 1 million neurons and 120 million synapses, and Intel has built a research cloud called the Intel Neuromorphic Research Community (INRC) that lets researchers access it. The chip has been used in experiments ranging from robotic arm control to olfactory sensing – training a chip to identify hazardous chemicals by "smell" – with results that showed significant efficiency gains over conventional approaches.
IBM's NorthPole chip, announced in late 2023, takes a somewhat different approach. It's not a pure neuromorphic chip in the spiking sense, but it's deeply inspired by brain architecture – eliminating the separation between memory and compute to dramatically reduce data movement. IBM's published benchmarks showed NorthPole running image recognition tasks at 25 times the energy efficiency of comparable GPU-based systems at the same process node.
Intel also operates Hala Point, the world's largest neuromorphic system as of 2024, built from 1,152 Loihi 2 chips. It contains about 1.15 billion neurons – comparable in scale, though not in capability, to a portion of a human brain – and is being used for research into sustainable AI computing.
BrainScaleS and SpiNNaker are two other research-focused neuromorphic systems developed in Europe under the Human Brain Project, both exploring different physical implementations of brain-inspired computing at scale.
Where Neuromorphic Chips Actually Have an Advantage
It's worth being clear about what neuromorphic chips are currently good at and what they're not. They're not about to replace GPUs for training large language models or generating images. Those workloads are dense, synchronous, and matrix-heavy – exactly the profile that GPUs are optimized for.
Where neuromorphic chips show real promise is in edge computing, real-time sensing, and low-power inference. Edge computing refers to processing that happens on the device itself rather than in the cloud – in a smartphone, a sensor, a robot, a wearable. These contexts have strict power budgets and often involve continuous streams of sensory data that don't need dense computation, just fast and efficient pattern recognition.
Consider a security camera doing real-time object detection, or a hearing aid that continuously processes audio to filter out background noise, or a drone navigating in real time using visual cues. All of these involve event-driven sensory input where most of the data stream is "nothing interesting is happening" – exactly the kind of sparse, low-activity processing that neuromorphic chips handle efficiently. Running these tasks on a conventional chip means powering all that computation continuously. Running them on a neuromorphic chip means the hardware is largely idle until something worth processing arrives.
There's also promising work in robotics, where neuromorphic chips can process proprioceptive feedback (information about the body's own position and movement) in real time with very low latency and power use – something that's notoriously difficult to do efficiently with conventional hardware.
The Honest Challenges
Neuromorphic computing has been "the future" for a long time, and there are real reasons it hasn't replaced conventional hardware. The most significant is programmability. Writing software for spiking neural networks is fundamentally different from the deep learning frameworks – PyTorch, TensorFlow, JAX – that millions of developers already know. The tooling is less mature, the talent pool is smaller, and porting existing models to run on neuromorphic hardware is often more art than science.
Spiking neural networks are also harder to train effectively. Conventional neural networks are trained using backpropagation, a well-understood algorithm that works cleanly on differentiable functions. Spiking networks use discrete, non-differentiable spikes, which makes gradient-based training methods more complicated to apply. Researchers have developed workarounds, but the training process is still less efficient and less reliable than what's possible with conventional networks.
There's also the question of what tasks neuromorphic hardware is actually needed for. As GPU efficiency continues to improve and specialized inference chips become more common, the efficiency gap between conventional and neuromorphic hardware has narrowed in some areas. The question isn't just "can neuromorphic chips do this more efficiently?" but "efficiently enough to justify the engineering cost of switching?"
Why This Matters Beyond the Lab
The energy question is the reason neuromorphic computing is gaining renewed urgency. As AI workloads scale – more inference requests, more edge deployments, more continuous sensing – the power consumption of conventional hardware becomes a genuine constraint. Data centers already account for a significant and growing share of global electricity consumption, and inference (running AI models, not just training them) is increasingly where that cost lives.
Neuromorphic chips won't solve this problem alone, but they represent a legitimate architectural path toward AI hardware that does more with less. Intel estimates that neuromorphic approaches could eventually enable certain workloads to run at 1,000 times the energy efficiency of today's GPU-based systems. Even a fraction of that gain, applied at scale, would be meaningful.
There's also a longer-term research implication. Neuromorphic hardware is one of the best tools available for studying and simulating biological neural circuits, which has value well beyond commercial AI – for neuroscience, for understanding how intelligence works, and for developing AI architectures that are genuinely different from the transformer-based models that currently dominate the field.
What's Coming Next
The trajectory is toward integration rather than replacement. It's likely that future AI systems – especially those deployed at the edge – will use heterogeneous hardware: a conventional processor for dense computation, a neuromorphic co-processor for event-driven sensing and inference, and specialized accelerators for whatever else the workload demands. Each component handling what it's best at.
Intel has signaled continued investment in Loihi and Hala Point. Research groups across Europe, the US, and China are publishing regularly on advances in both spiking neural network training and physical chip design. And as the energy costs of conventional AI hardware become harder to ignore, the appetite for alternatives is growing.
Neuromorphic chips aren't a revolution that's about to happen overnight. But they're a serious and increasingly well-funded part of the answer to a question the industry genuinely needs to solve: how do you keep scaling AI without the energy cost scaling proportionally?
FAQ
Are neuromorphic chips available to consumers today? Not in consumer products yet. Current neuromorphic chips are primarily research tools available through academic programs or research partnerships. Intel's INRC gives researchers cloud-based access to Loihi 2, but there's no consumer hardware running neuromorphic chips as of 2024.
Is neuromorphic computing the same as quantum computing? No – they're completely different approaches. Quantum computing uses quantum mechanical phenomena (superposition, entanglement) to perform certain types of computation exponentially faster. Neuromorphic computing is classical computing that mimics the architecture of biological brains to improve energy efficiency and real-time processing. They address different problems and aren't in competition with each other.
Can existing AI models run on neuromorphic chips? Not directly. Models trained using conventional deep learning frameworks need to be converted or retrained to run on neuromorphic hardware. Some tools and methods exist for this conversion, but it's not a straightforward port – it often requires significant engineering work and sometimes a redesign of the model architecture.
What's the difference between neuromorphic chips and regular AI chips like TPUs? TPUs (and similar chips) are conventional von Neumann processors optimized for the matrix multiplication operations used in deep learning. They're synchronous, power-hungry at scale, and excellent at dense computation. Neuromorphic chips are event-driven, co-locate memory and processing, and excel at sparse, low-power computation. They're solving different problems with different architectural approaches.
The Bigger Picture
Most of the conversation about AI hardware focuses on making conventional chips faster – more cores, better memory bandwidth, higher throughput. Neuromorphic computing asks a more fundamental question: what if the architecture itself is the wrong starting point? That's a harder problem to solve, but potentially a more important one.
The brain processes the world continuously, at roughly 20 watts. The hardware we've built to imitate what it does requires megawatts. Closing that gap is one of the more interesting engineering challenges in technology right now – and neuromorphic chips are one of the more credible paths toward doing it.
📚 Sources
Davies, M. et al. (2018). Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro – https://ieeexplore.ieee.org/document/8259423
Intel Newsroom. (2024). Intel's Hala Point: World's largest neuromorphic system – https://www.intel.com/content/www/us/en/newsroom/news/intel-builds-worlds-largest-neuromorphic-system.html
Modha, D. et al. (2023). Neural inference at the frontier of energy, space, and time. Science – https://www.science.org/doi/10.1126/science.adh1174
Schuman, C. D. et al. (2022). Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science – https://www.nature.com/articles/s43588-021-00184-y
Markovic, D., Mizrahi, A., Querlioz, D., & Grollier, J. (2020). Physics for neuromorphic computing. Nature Reviews Physics – https://www.nature.com/articles/s42254-020-0208-2
IEA. (2024). Electricity 2024: Analysis and forecast to 2026. International Energy Agency – https://www.iea.org/reports/electricity-2024
Intel Labs. Loihi 2 neuromorphic chip. Intel Research – https://www.intel.com/content/www/us/en/research/neuromorphic-computing.html
