Neuromorphic Computing

The brain runs on about 20 watts; a single NVIDIA H100 accelerator is rated around 700 watts. Neuromorphic computing asks whether silicon should stop pretending cognition is only matrix multiplication and start exploiting spikes, locality, silence, and time. The bet is not that chips should copy biology perfectly. The bet is that the memory-compute split is expensive, and brains are proof that useful computation can happen without it.

How it works

A normal computer keeps memory and processing separate. Data moves from DRAM to cache to arithmetic units and back again, burning energy on transport as much as logic. Neuromorphic chips move the weights and state closer to the compute, then fire only when something changes.

The core unit is usually a spiking neuron. It accumulates input, crosses a threshold, emits a short pulse, then resets or adapts. That makes time part of the computation rather than a nuisance to be sampled at 30 or 60 frames per second. An event camera looking at a still room may send almost nothing; a frame camera still ships the whole image again.

Biology gives the target scale: roughly 86 billion neurons, around 100 trillion synapses, and a resting power draw near 20 W. Current chips are nowhere near that full package, but they can win on narrow jobs where sparse events matter more than dense arithmetic.

Hardware landmarks

System	Date	What matters
Carver Mead's neuromorphic framing	1990	Put analog VLSI and nervous-system computation into one research program
IBM TrueNorth	2014	1 million programmable neurons and 256 million synapses on a 70 mW research chip
Intel Loihi	2017	On-chip learning and asynchronous spiking for research workloads
Intel Loihi 2	2021	Up to 1 million neurons per chip, a better programming stack, and more flexible neuron models
IBM NorthPole	2023	Memory placed next to compute for neural inference; not a brain copy, but an attack on data movement
Intel Hala Point	2024	1,152 Loihi 2 processors, about 1.15 billion neurons, installed at Sandia National Laboratories
Loihi 3 claims	2026	to verify: reported 4 nm chip with higher density and graded spikes

NorthPole is a useful warning against sloppy language. It is often grouped with neuromorphic work because it attacks the same energy bottleneck, but it is not primarily a spiking-neuron machine. Its Science paper reported 25x better energy metric than a 12 nm GPU and 5x better than a 4 nm GPU on tested inference workloads. That is not a general AI victory; it is a specific architectural result.

Where it shows up

Neuromorphic hardware fits best where the world is already event-driven: vision, touch, hearing, anomaly detection, motor control, and always-on sensing. A robot avoiding a falling object does not need a polished caption. It needs a few milliseconds of useful signal without heating its battery pack.

The cleanest pairing is event-based vision. Dynamic Vision Sensor cameras emit pixel changes rather than full frames, so a chip can process motion without rereading a whole image. That makes concept temporal sparsity the real prize: if nothing changes, almost nothing fires.

The less proven target is language. Transformers won because dense linear algebra maps cleanly onto GPUs and training recipes are mature. Spiking large language models exist as demos and papers, but the gap between "can imitate a trained network" and "can train and serve frontier models cheaply" is still large.

What's contested

The contested question is not whether neuromorphic chips can be efficient. They can. The question is whether they can be useful outside the workloads selected to flatter them.

Benchmarks are messy because GPUs, CPUs, FPGAs, event cameras, and neuromorphic chips often run different models under different accuracy targets. A 100x energy claim means little unless latency, accuracy, batch size, sensor type, and training cost are pinned down. The field still lacks a benchmark suite with the social force that ImageNet had for vision or MLPerf has for conventional inference.

The deeper unknown is training. Backpropagation expects smooth functions; spikes are discontinuous events. Surrogate gradients, spike-timing-dependent plasticity, and hybrid ANN-to-SNN conversion all work in pieces, but none has become the default recipe for broad machine intelligence.

Why this has to do with other realms

Neuromorphic computing is biology smuggled into hardware design. concept distributed cognition matters because octopus arms, insect ganglia, and the enteric nervous system show that control does not need one central clock. Computation can live at the edge of the body, close to the sensor and actuator.

It also belongs with physics. If brains operate near critical regimes, then the interesting chip may not be the most stable one. concept brain turbulence asks whether useful cognition comes from dynamics balanced between silence and runaway cascade.

Space gives the harshest test. A gram-scale probe in mission breakthrough starshot cannot carry a data center, cannot wait 4 years for instructions, and cannot waste watts on empty frames. If neuromorphic computing has a killer environment, it may be a spacecraft that must notice rare events while mostly staying quiet.

An open question

Can neuromorphic chips find their ImageNet moment: one public task, one hard metric, and one result that makes GPUs look like the wrong tool?

Key Sources

Carver Mead, "Neuromorphic Electronic Systems" (Proceedings of the IEEE, 1990) - the founding architecture argument.
Paul A. Merolla et al., "A Million Spiking-Neuron Integrated Circuit with a Scalable Communication Network and Interface" (Science, 2014) - IBM TrueNorth.
Mike Davies et al., "Loihi: A Neuromorphic Manycore Processor with On-Chip Learning" (IEEE Micro, 2018) - Intel's first Loihi paper.
Dharmendra S. Modha et al., "Neural inference at the frontier of energy, space, and time" (Science, 2023) - IBM NorthPole.
Intel, Hala Point neuromorphic system announcement (2024) - to verify exact installed configuration and performance claims.
to verify: Intel Loihi 3 public technical brief, 2026.

Abhishek