How does a photonic tensor core work?

The news coverage dont explain. It is clear from the paper. There are multiple units that perform the dot-product operation, using Wavelength division multiplexing. These wavelengths, which represent the entries of a 4x1 vector (modulated carriers), are weighted through multi-bit photonic memories that are written electrically. The weighted wavelengths are summed together by photodetector. And this happens for all the 16 dot-products of a 4x4 matrix-matrix multiplication. very similar to electronic TPU operations, with difference that most operations are performed while photons travel in the waveguides. It is elegant. no separation between memory and processing.

What Anelov says seems correct. https://arstechnica.com/science/2018/07/neural-network-implemented-with-light-instead-of-electrons/ reads as a diffractive optics network with 3d printed "boards". It can not be tuned. You need to 3d print new modules for new neural networks. The photonic tensor core uses micron size waveguides, so much smaller and it seems that the memory can be written and it can potentially implement multiple neural networks.