Low-power AI at the edge: UOC team advances spiking neural networks for real workloads
AI runs on electricity. The International Energy Agency estimates data centres already draw roughly 1.5% of global electricity, with demand set to double by 2030 if trends hold. For engineering teams, that's a scaling bottleneck-and a cost line you can't ignore.
Researchers at the Universitat Oberta de Catalunya (UOC) share two studies that push a clear alternative: spiking neural networks (SNNs) built for low-power, high-performance deployment on accessible hardware. The goal is straightforward-treat energy as a first-class design constraint so teams can ship AI that's sustainable, affordable, and resilient in weak-connectivity environments.
Why this matters for IT and development teams
Energy-aware AI reduces infrastructure spend, heat, and latency. It also enables on-device inference for sensors, robots, and vehicles-without relying on GPUs or persistent cloud links. That's better for privacy, better for uptime, and better for budgets.
Study 1: Distributed SNNs on low-cost edge hardware
The first study shows SNNs running on off-the-shelf components-Raspberry Pi 5 paired with a BrainChip Akida accelerator-delivering high performance while staying under 10 watts. The workflow covers training, conversion, and deployment without a data centre or GPU.
Devices coordinate using protocols you already know: MQTT (Message Queue Telemetry Transport), SSH, and Vehicle-to-Everything (V2X). Results can be shared in under a millisecond with an energy cost of about 10-30 microjoules per operation. Think distributed, low-latency AI for transport, environmental monitoring, and industrial IoT.
This approach also improves accessibility and privacy. Schools, hospitals, rural deployments, and resource-constrained teams can run meaningful AI locally-safely and sustainably.
Study 2: Energy-aware autonomous driving with SNNs
The second study compares SNNs to standard convolutional networks on tasks like steering angle prediction and obstacle detection. With the right encoding, SNNs deliver a strong accuracy/energy balance while using 10-20x less energy than CNNs.
The team also proposes a practical metric that blends accuracy with consumption, helping engineers choose models based on real deployment costs instead of benchmark scores alone.
What to do next (practical steps)
- Treat energy as a KPI. Track watts, joules/inference, and latency alongside accuracy.
- Prototype a low-power stack: Raspberry Pi 5 + Akida for SNN acceleration, with containerized services.
- Use MQTT for lightweight messaging; secure sessions with SSH; plan for V2X where mobility matters.
- Build an edge-first pipeline: local preprocessing, on-device inference, selective uplink of summaries.
- Adopt an efficiency metric in model selection-opt for the best accuracy per joule, not just per FLOP.
- Target deployments where privacy and resilience matter: clinics, schools, factories, rural sites.
Where this fits
Both studies contribute to practical, energy-thrifty AI systems that are easier to deploy and maintain. They align with UN SDGs 9 (Industry, Innovation and Infrastructure), 11 (Sustainable Cities and Communities), and 13 (Climate Action)-but the benefit is immediate for engineering teams: less power, less heat, lower cost, same business impact.
Further reading
- IEA: Data centres and network energy use
- Eco-Efficient Deployment of Spiking Neural Networks on Low-Cost Edge Hardware (IEEE Networking Letters)
Your membership also unlocks:
