In case anyone is wondering, from looking at their patents they are doing fairly standard Fourier optics. That is, one lens takes a Fourier transform of an image onto a spatial light modulator, then projects it back onto a detector chip. This amounts to a convolution operation in real-space. Since all the optical stuff happens in parallel, it can be performed as quickly as the modulators can be set up and the detectors read out (which can also be parallelized.)

The classic book about this kind of approach is Goodman: https://www.amazon.com/Introduction-Fourier-Optics-Joseph-Go...

It's basically a hardware accelerator for a convolution - an important step in a neural network, but it isn't a general purpose processor (so beware their benchmarks of "Ops per second" - these aren't equivalent to CPU/GPU ops.)