Scientists from Florida University jointly with their colleagues from Florida Semiconductors Institute, Californian University in Los Angeles and George Washington University developed a silicon chip using light for performing calculations in AI systems. The new technology allows for many-fold reduction of energy consumption and for speeding-up data processing without compromising accuracy.
Modern artificial intellect models are becoming more and more complicated and require more and more energy resources. Colossal masses of energy are needed to train the models and then – to assure their operation. The convolution operation is especially energy-consuming, and this is the key tool of neural networks allowing for identifying the regularities and consistent patterns in images, videos and texts. The biggest amounts of energy and time are consumed at this particular stage.
To solve this issue the researchers turned to photonics — the technology using light for data transfer. In the traditional chips the calculations are done driven by the movement of electrons, which is inevitably accompanied by energy losses and heat release, hence, requires additional cooling. In the new system some operations are performed by laser radiation: photons do not heat the chip to the same extent as electrons, and practically do not consume extra energy. To implement this principle, the scientists put miniscule Fresnel lenses, which are thinner than a human hair, on the surface of the chip. These ultra-thin optical elements convert data into laser pulses and allow for performing mathematical operations with minimum energy consumption.
In the course of the tests the prototype demonstrated impressive results: it classified hand-written figures with 98% accuracy — at the level of traditional systems, but consuming significantly less energy. Moreover, the chip is capable to process several data flows simultaneously using lasers of different colors. This approach, known as wavelength division multiplexing, opens the way to further performance improvement.
According to Professor Volker Zorger heading the project, this is for the first time that the researchers succeeded in placing such optical calculations directly on the chip and apply them in neural network. The scientists are convinced that in future such solutions may become standard: the manufacturers including NVIDIA are already using certain optical elements in their systems. In the long-run, it will simplify the integration of new technologies and bring the creation of fully optical calculations closer.
