A team of MIT scientists and their international colleagues have created what they claim is the first all-photonic processor for artificial intelligence applications. The photonic processor performs no worse than its analogues on silicon transistors, but carries out calculations with much less energy consumption. This is especially important for the creation of “thinking” peripherals – lidars, cameras, communication devices and others, to which a direct road is now open.
Image source: Sampson Wilcox, Research Laboratory of Electronics
The main challenge in creating an all-photonic chip for AI is that light is good at linear computation, while nonlinear computation requires significant energy consumption. To carry out the latter, special blocks are needed, because photons react with each other only under special conditions. Therefore, previously linear operations, for example, matrix multiplication, were carried out by a photonic unit, and for nonlinear calculations, the light signal was converted into the form of an electrical pulse and then processed the old fashioned way – by a conventional processor made of silicon transistors.
Scientists from MIT set themselves the goal of creating a single processor that would have a light signal at the input and a light signal at the output without the use of silicon coprocessors. According to them, using previous work and the findings of foreign colleagues, they achieved their goal.
The optical device the researchers developed was able to perform key calculations for a classification task using machine learning in less than half a nanosecond, while achieving an accuracy of more than 92%—performance that is on par with traditional equipment. The created chip consists of interconnected modules that form an optical neural network and is manufactured using commercial lithographic processes, which can ensure scalability of the technology and its integration into modern electronics.
Scientists have gotten around the problem with nonlinear photonic calculations in an interesting way. They developed the NOFU unit, a nonlinear optical functional unit integrated into the optical processor, which made it possible to use electronic circuits along with optical circuits, but without switching to external operations. Apparently, the NOFU block was chosen as a compromise between purely photonic nonlinear circuits and classical, electronic ones.
First, the system encodes the parameters of the deep neural network in light pulses. An array of programmable beam splitters then performs matrix multiplication of the input data. The data is then transferred to the NOFU programmable layer, where nonlinear functions are implemented by transmitting light signals to photodiodes. The latter, in turn, translate the light signal into electrical impulses. Since this stage does not require external amplification, NOFU units consume very little power.
«We remain in the optical region all the time, until the end, when we want to read the answer. This allows us to achieve ultra-low latency,” say the study authors.