How large can an optical matrix-matrix multiplier be constructed?

In this talk, based on the key operations governing all these matrix-matrix multipliers, we propose a statistical model to quantitatively analyze the numerical accuracy that these multipliers could deliver. Our study indicates that no matter what particular scheme is used, statistically the numerical accuracy limit caused by the use of analog computation mechanism is much more serious than the dynamic range and other limits to a matrix-matrix multiplier. Our study also shows that when these multipliers are used for, instead of generating algebraic results, the single threshold operations, such as the operations required to implement a programmable logic array, a content- addressable memory, and a neural network, a better accuracy can be expected depending on the used threshold positions.