Neuromorphic Computing

The human brain processes complex sensory input, learns from sparse feedback, and navigates uncertain environments on 20 watts—less than a Wi-Fi router. In 2024, a single Google search consumed roughly 0.3 watt-hours, meaning the brain could run 240,000 searches per day within its total power budget. Neuromorphic chips attempt to close that gap not by mimicking brain function in software, but by rebuilding the physics of computation in silicon: no clock, no fetch-execute cycle, and—critically—no power draw when idle.

How it works

Biological neurons communicate via spikes, all-or-nothing electrical pulses that travel across synapses only when input signals cross a threshold. Between spikes, neurons consume near-zero energy. Neuromorphic chips replicate this event-driven behavior: transistors act as artificial neurons, and capacitors simulate synaptic integration. Computation occurs only when an input event (e.g., a pixel change in a camera) triggers a spike.

Unlike von Neumann machines (1945), where memory and processing units are separate, neuromorphic architectures embed memory in the synapse—each connection stores its own weight locally. There is no bus to saturate, no pipeline to stall. The chip’s behavior emerges from decentralized, asynchronous dynamics.

Two principles dominate:

Sparsity: At any instant, only 1–10% of artificial neurons fire—mirroring cortical activity in mammals.
Temporal coding: Information is encoded in the timing of spikes, not their amplitude. A millisecond delay can flip a synaptic weight via spike-timing-dependent plasticity (STDP), a hardware-level implementation of Hebbian learning.

Specific examples

Intel Loihi 2 (2021)

1 million neurons, 120 million synapses per chip
128 neuromorphic cores, 14-bit precision
On-chip learning via programmable STDP
100× more energy-efficient than GPU for control tasks (e.g., robotic balancing)

Hala Point (2024)

1,152 Loihi 2 chips at Sandia National Labs
1.15 billion neurons, 128 billion synapses
140,544 cores, 2,600-watt draw
Size: 6U rack, comparable to a microwave
Throughput: 20 petaops/sec at >15 TOPS/W efficiency
Roughly matches the neuron count of a capuchin monkey cortex

IBM NorthPole (2023)

Hybrid design: drops spiking, but fuses memory and logic
256 MB on-chip SRAM, zero off-chip memory access
22× faster, 25× more efficient than GPU on ResNet-50
No clock: computation flows as data arrives

Intel Loihi 3 (2026, announced)

4 nm process, 8 million neurons per chip
“Graded spikes” (32-bit intensity) bridge SNN efficiency with DNN precision
Targets cloud-edge parity: on-device inference matching datacenter models

What's contested

We assume the brain’s efficiency comes from sparsity and event-driven processing—but we may be wrong. The human brain sustains 86 billion neurons on 20 watts. Hala Point uses 2,600 watts for 1.15 billion neurons: a 100-fold deficit in neurons-per-watt. This gap suggests we’ve copied the architecture without capturing the substrate physics: ion channels, myelin sheaths, glial support, and 3D volumetric wiring all contribute to biological efficiency. Are spikes even the right abstraction? Some neuroscientists argue sub-threshold analog signaling, dendritic computation, or astrocyte modulation do most of the work—none of which SNNs model.

Another open question: Can SNNs scale to general intelligence? Current benchmarks show gains in latency and power for narrow tasks—robot locomotion, gesture recognition, anomaly detection—but no SNN has matched a transformer on language or reasoning. The field lacks a clear path from efficient perception to abstract thought.

Why this has to do with other realms

dest proxima centauri is 4.24 light-years away. To reach it in 50 years requires a propulsion system we don’t have—or an AI small and efficient enough to fit on a gram-scale probe. Breakthrough Starshot aims to launch thousands of solar-sail nanocrafts at 20% lightspeed. At that scale, power budgets are measured in milliwatts. A GPU is impossible. A von Neumann CPU would overheat. But a neuromorphic controller—one that only wakes when a star drifts into frame or a dust particle strikes the sail—could survive the trip. The brain’s physics may not just inspire AI; it may be the only physics that fits interstellar flight.

An open question

If we replicated the brain’s wiring diagram exactly in silicon, but powered it with a wall outlet, would it think faster—or just overheat?

Key sources

Mead, C. (1990) Neuromorphic Electronic Systems — original formulation of the term and concept
Davies, M. et al. (2018) Loihi: A Neuromorphic Manycore Processor (IEEE Micro) — Loihi 1 design
Akopyan, F. et al. (2015) TrueNorth: Design and Tool Flow of a 65 mW 1 Million Neuron Programmable Neurosynaptic Chip (IEEE TCSVT) — IBM’s landmark chip
Furber, S. (2016) Large-scale neural models: opportunities and challenges (Phil. Trans. R. Soc. A) — critical perspective on biological fidelity
to verify: NorthPole benchmark claims (IBM Research, 2023 press release)

Abhishek