# [TECH-Idea] Neuromorphic Computing
⏎
**Neuromorphic computing** uses chip architectures inspired by the brain's neural structure — sparse, event-driven spiking neurons, co-located memory and compute, asynchronous operation — achieving orders-of-magnitude better energy efficiency than conventional GPUs for AI inference and edge computing.
⏎
## Overview
⏎
Training GPT-4 consumed an estimated USD 100M in compute and ~50 GWh of electricity. As AI inference scales to billions of users, energy consumption becomes a critical constraint: without efficiency gains, AI could consume 10%+ of global electricity by 2030. Neuromorphic chips address this by abandoning the von Neumann bottleneck (separating memory from compute) and the synchronous clock (wasting energy on idle cycles).
⏎
**Leading systems**:
- **Intel Loihi 2** (2021): 1 million neurons, 120 million synapses; 1,000× better energy efficiency than conventional hardware for sparse event-driven tasks (gesture recognition, anomaly detection).
- **Intel Hala Point** (2024): 1.15 billion neurons (largest neuromorphic system), energy consumption equivalent to a lightbulb while running complex AI.
- **IBM NorthPole** (2023): brings all model parameters on-chip, eliminating the memory bandwidth bottleneck; 22× more energy efficient than GPU for inference.
- **BrainScaleS** (Heidelberg): 512 wafer-scale neuromorphic processors; 10,000× real-time acceleration of neural simulation.
- **SpiNNaker2** (Manchester/Dresden): 10 million neurons, brain-scale simulation.
⏎
Key insight: spiking neural networks (SNNs) communicate binary spikes only when information changes — analogous to biological neurons. This sparse, event-driven communication eliminates 90%+ of compute cycles vs. dense matrix multiplication in conventional transformers.
⏎
## Key Actors
⏎
Intel (Loihi, Hala Point), IBM (NorthPole, TrueNorth), Qualcomm (NPE neuromorphic DSP), SpiNNaker (Manchester), BrainScaleS (Heidelberg), Airliquide (industrial applications), Innatera (embedded neuromorphic), Prophesee (event-based vision sensors).
⏎
## Economic Value
⏎
AI inference hardware market: USD 150B+/year by 2030. Electricity savings: USD 30B+/year if neuromorphic replaces GPU inference for always-on AI. Edge AI (autonomous vehicles, robotics, IoT, wearables) market: USD 50B+/year by 2030, where battery life makes neuromorphic energy efficiency essential. Total addressable market: USD 200B+/year by 2030.
⏎
## Notes
⏎
The term and concept were coined by Carver Mead (Caltech) in 1990; it took 30 years and the AI energy crisis to make neuromorphic commercially urgent.
⏎
## Discovery Character
⏎
**Surprise level**: Moderate — the concept was known for 35 years; the surprise is how quickly energy constraints are forcing the industry to revisit it. IBM's NorthPole achieving 22× efficiency gains over GPUs in 2023 surprised the AI hardware community.
⏎
**Mode**: Systematic over 30 years; now rapidly becoming commercially systematic under the pressure of AI energy costs. The parallel with how energy crises historically drive efficiency innovation (OPEC 1973 → fuel injection, LED → solid-state lighting) is striking.
⏎
## What This Enables
⏎
- **[TECH-Idea] Autonomous Robots & Physical AI** — edge-deployable neuromorphic chips enable always-on perception and fast reaction in mobile robots without GPU-scale power consumption.
⏎
# Parents
⏎
* [SCI] Deep Learning⏎