title: "What is neuromorphic hardware?" description: "Chips that mimic the human brain using silicon neurons and synapses — and why this could redefine the future of AI and energy efficiency." date: "2024-08-12" tags: ["hardware", "ai", "neuromorphic"] featured: true
Traditional computing follows a simple, brutally efficient model: the CPU fetches an instruction, memory delivers the data, repeat. But when the problem is to simulate the way we think, that model stalls — not because of speed, but because of shape. Neuromorphic hardware is the attempt to change the shape.
The term was coined in the 1980s by Carver Mead, who looked at the human brain and asked a direct question: what if we designed chips that worked more like neurons than like logic gates? Four decades later, that question is gaining traction — IBM, Intel, and research labs are building silicon that learns, processes in massive parallel, and burns a fraction of the energy.
The brain as architecture
To understand neuromorphic hardware, start with the model: the brain has around 86 billion neurons and a hundred trillion synapses, organized without a central clock. Each neuron fires only when the sum of incoming signals crosses a threshold. There is no instruction, no external memory, no bus. Computation and storage live in the same place.
Compare that to the von Neumann architecture running on your phone: CPU separated from RAM, synchronous clock cycles, and the famous memory wall — the bottleneck that shows up when processor and memory fight for bandwidth. Moving data between the two costs more energy than computing on it.
The human brain runs on ~20 watts. A data center training GPT-4 runs on megawatts. The difference isn't the number of operations — it's how they're organized.
How a neuromorphic chip works
A neuromorphic processor has no global clock. Instead, thousands of small cores — each simulating tens to hundreds of artificial neurons — exchange messages as spikes: discrete pulses fired only when there's relevant information. If nothing happens, nothing is processed. If everything happens, everything fires in parallel.
That changes three things:
- Local computation — memory and processing on the same silicon. No bus, no memory wall.
- Event-driven — only consumes energy when there's a signal. On sparse workloads, the savings are absurd.
- Plasticity — synapses adjust weights in real time, without offline training. The chip learns while it runs.
Who is building it
The three projects that show up in any serious conversation about this:
Intel Loihi 2
Announced in 2021, Loihi 2 packs up to 128 thousand artificial neurons and a million synapses per chip. Intel arranges Loihi arrays in systems like Hala Point — ~1.15 billion neurons in a single rack. Today's focus is robotics, optimization, and event-driven vision.
IBM TrueNorth and NorthPole
IBM started with TrueNorth (2014, ~1 million neurons) and in 2023 announced NorthPole — a chip that combines a neuromorphic approach with deep neural network inference. In vision benchmarks, NorthPole is up to 25× more energy-efficient than comparable GPUs.
SpiNNaker (University of Manchester)
An academic project that scaled to a million ARM cores running in parallel, simulating spiking neural networks. Less about absolute efficiency, more about studying the biological brain in real time.
Why it matters
Generative AI, the way it's built today, is expensive. Training large models consumes megawatt-hours; production inference is costly even when optimized. A big part of that bill comes from the architecture: GPUs are absurdly good at dense matrix multiplication, but terrible when the signal is sparse, asynchronous, or continuous in time.
Neuromorphic hardware shines in exactly those cases:
- Always-on sensors — wearables, IoT, drones — where battery life is decisive.
- Event-driven audio and vision processing — DVS cameras that only report pixel changes, not whole frames.
- Adaptive robotics — systems that need to learn in the field without going back to the data center to retrain.
- Edge AI — local inference on devices with tight energy budgets.
What still doesn't work
It would be dishonest not to say: the ecosystem is early. There's no mature tensorflow for neuromorphic hardware. Each chip has its own toolchain, its own modeling language, its own quirks. Training a spiking neural network that matches the accuracy of a traditional network is still art, not engineering.
Worse: not every problem maps well to this paradigm. Tasks that depend on dense floating-point arithmetic — like training large LLMs — will keep living on GPUs for a long time.
Why I write about this
I've been working in mobile for seven years. We spend our days optimizing battery, bandwidth, latency. When I read about Loihi 2 or NorthPole, what catches my attention isn't the brain-simulation part — it's the do-more-with-less-energy part. That's the mobile game, taken to the extreme.
The next decade of AI won't be won by larger models alone. It will be won, in part, by different hardware. Worth paying attention.