Direct tensor processing with coherent light
7 days ago
- #tensor processing
- #neural networks
- #optical computing
- Tensor processing is crucial for modern technologies like AI and data analytics, but current optical methods struggle with tensor-based tasks.
- Parallel Optical Matrix-Matrix Multiplication (POMMM) enables fully parallel tensor processing through single coherent light propagation, improving computational power and efficiency.
- POMMM demonstrates high consistency with GPU-based matrix-matrix multiplication in both real-valued and complex-valued domains.
- The approach supports various tensor processing applications, including convolutional and vision transformer neural networks, showcasing adaptability and scalability.
- POMMM leverages multi-wavelength and large-scale expansion, offering a scalable, high-efficiency foundation for next-generation optical computing.
- Experimental validation shows POMMM's reliability and accuracy, with low error rates compared to GPU-based methods.
- POMMM's theoretical compatibility and efficiency outperform existing optical computing paradigms, highlighting its transformative potential.