Neuromorphic Computing

last updated 2026-06-09 · +8 sources in last 30d

Physics / mechanism

Neuromorphic computing implements neural computation in hardware by co-locating memory and processing—eliminating the von Neumann bottleneck. Core primitive is the artificial neuron/synapse: spiking neural networks (SNNs) communicate via sparse binary events (spikes) rather than continuous activations, dramatically reducing switched capacitance. Key substrates: CMOS (Intel Loihi 2: 1M neurons, 120M synapses, 10× better energy/inference than GPU on sparse workloads), memristive crossbars (PCM, RRAM, OTS selectors), and emerging ferroelectric FETs. Synaptic density, spike encoding efficiency, and on-chip learning (STDP) are the differentiated parameters. Energy figures: 10–100 pJ/synaptic operation vs. ~1 nJ on GPU. Still pre-productisation: no dominant ISA, fragmented toolchains.

Competitive landscape

Competing inference approaches: standard GPU/TPU inference (mature, high throughput, energy-hungry), analog in-memory computing (similar von Neumann escape, continuous weights, noise-limited precision), and photonic neural networks (ultralow latency, WDM parallelism, limited nonlinearity). SNNs compete directly with quantised transformers on edge inference benchmarks. The clearest competitive wedge is always-on sensory processing: radar, LiDAR, event cameras—domains where sparse, asynchronous data maps naturally to spike codes and where GPU idle power is prohibitive.

ApproachEnergy/opProgrammabilityMaturity
Neuromorphic (SNN)~10–100 pJLow-mediumTRL 4–6
Analog IMC~1–10 pJLowTRL 4–5
Digital GPU/TPU~1 nJHighProduction

Where value accrues (vehicle-agnostic)

Neuromorphic’s economic prize is the always-on sensing/inference tier inside devices that cannot afford a GPU’s power budget — captured today mostly by the SoC platform owners and the foundries supplying low-leakage process + embedded NVM, not by standalone neuromorphic vendors (the category’s base rate is poor; see Low Power Edge Compute). The credible substrate is a GF-shaped node: 22FDX / 12LP (low leakage, embedded NVM incl. eMRAM, FD-SOI body-bias) suits the always-on, low-VDD requirement.

Companies using

Connected ideas

Sources

Frontier (open questions)

Recent mentions

Frontier questions