Vellex offers a physics‑based analog intelligence accelerator that enables on‑device AI model training for edge hardware such as satellites, drones, wearables, and IoT sensors. By mapping optimization calculations onto physical systems, the accelerator converges models in milliseconds while consuming under 10 mW, eliminating the need for power‑hungry GPUs and cloud‑based data transfers. This hardware solution provides up to 17,000× faster training, preserving battery life, reducing latency, and keeping data private for autonomous edge applications.
Funding
Funding not disclosed
Founders
Product
Problem
Edge devices such as satellites, drones, wearables, and IoT sensors cannot perform on‑device AI model training because conventional digital AI hardware requires watts of power and relies on frequent cloud communication, creating latency, high bandwidth costs, and security concerns.
Solution
Vellex provides a physics‑based analog intelligence accelerator that maps optimization calculations onto physical systems, allowing AI models to converge in milliseconds with power consumption below 10 mW—less than a typical Bluetooth radio. The architecture eliminates the need for iterative digital computation, enabling real‑time, on‑device training without cloud off‑loading. By drastically reducing training time (up to 17,000× faster) and energy use, the platform lets autonomous edge systems continuously adapt to new data while preserving battery life and data privacy. Vellex’s solution is delivered as a hardware accelerator that can be integrated into existing edge platforms, supporting a range of applications from satellite remote sensing to wearable personalization.
Target Audience
Primary customers are hardware manufacturers and system integrators developing autonomous edge solutions for satellite remote sensing, aerospace, robotics, drones, and wearable or hearable devices that require on‑device AI adaptation.
Features
- Analog compute engine that solves AI optimization problems via physical settling, removing millions of digital iteration steps
- Sub‑10 mW power envelope, enabling continuous learning on battery‑operated edge hardware
- Training latency reduced to milliseconds, delivering up to 17,000× faster model convergence compared to conventional GPUs
- Cloud‑independent operation that keeps raw data on‑device, mitigating bandwidth costs and security risks
- Proven performance benchmarks from Stanford University and Berkeley Lab
- Compatible with satellite imaging, drone navigation, robotics control, and wearable/hearable personalization use cases