Physics as Computation

Our PNN technology replaces traditional digital processors with electromagnetic metasurfaces, enabling AI computation through physics for breakthrough power efficiency.

How It Works

Physical Neural Networks leverage electromagnetic wave interactions to perform neural network computations in the analog domain, dramatically reducing power consumption while maintaining accuracy.

End-to-End Differentiable Training
Incremental Learning Support
Context-Aware Adaptation
1

Data to Signal

2

Physical Computation

3

Signal Captured

4

AI Result

See It In Action

A hardware-in-the-loop run of our RFNN-LLM pipeline: a 354M-parameter GPT-2 medium language model with its MLP layers offloaded onto our reconfigurable metasurface. Tokens are encoded into RF signals, transformed by electromagnetic waves on the RIS, and read back as the model's output — computed in real time by physics.

Our RFNN Device