EP 226 - Neuromorphic for LLMs on the Edge

In this episode, we spoke with Sean Hehir, CEO, and Jonathan Tapson, Chief Development Officer, of BrainChip about neuromorphic computing for edge AI. We covered why event-based processing and sparsity let devices skip 99% of useless sensor data, why joules per inference is a more honest metric than TOPS, how PPA (power, performance, area) guides on-device design, and what it will take to run a compact billion-parameter LLM entirely on device.

We also discussed practical use cases like seizure-prediction eyewear, drones for beach safety, and efficiency upgrades in vehicles, plus BrainChip's adoption path via MetaTF and its IP-licensing business model.

Key insights:

• Neuromorphic efficiency. Event-based compute minimizes data transfer and optimizes for joules per inference, enabling low-power, real-time applications in medical, defense, industrial IoT, and automotive.

• LLMs at the edge. Compact silicon and state-based designs are pushing billion-parameter models onto devices, achieving useful performance at much lower power.

• Adoption is designed to be straightforward. Models built in standard frameworks can be mapped to BrainChip's Akida platform using MetaTF, with PPA guiding silicon optimization and early evaluation possible through simulation and dev kits.

• Compelling use cases. Examples include seizure-prediction smart glasses aiming for all-day battery life in a tiny form factor and drones scanning beaches for distressed swimmers. Most current engagements are pure on-edge, with hybrid edge-plus-cloud possible when appropriate.

IoT ONE database: https://www.iotone.com/case-studies

The Industrial IoT Spotlight podcast is produced by Asia Growth Partners (AGP): https://asiagrowthpartners.com/