A startup company called Optalysis is trying to invent a fully-optical computer that would be aimed at many of the same tasks for which GPUs are currently used. Optalysis is claiming that they can create an optical solver supercomputer astonishing 17 exaFLOPS machine by 2020.
40 gigaFlops demonstration uses 500×500 pixels working 20 times per second. Each pixel does the work of about 8000 floating point operations in each of the cycles. Speeding up 427 million times to 17.1 exaFLOPS can be done with 500,000 X 500,000 pixels and 8550 cycles per second. They can use multiple LCD displays.
We can project out the minimum pixel size with visible light (without using superlensing and metamaterials to get to sub-wavelengths). The visible light spectrum are wavelengths from about 390 to 700 nm. So one micron one a side pixels would work for visible light. One million by one million pixels would then be one square meter.
Physical Review Letters had an article on Nanosecond Electro-Optic Switching of a Liquid Crystal in 2013. Using these type of crystals would enable 30 million cycles per second for the optical computer.
This extrapolation assumes that the Optalysis optical computing system is successful and looks at merging this some physical limits in with what Optalysis is doing. Ray Kurzweil has made famous projections mainly based on extrapolating the rate of improvement of computer systems over many decades. Here my extrapolation is based upon physical limits of a particular computational approach.
1,000,000 X 1,000,000 pixels by 30 million cycles by 8000 floating point operations equivalents per cycle would be 240 X 10^21 FLOPS [240 zettaFLOPS].
The 2030 supercomputer version of this could use a million of the desktop systems at the same time.