An international team of researchers from Germany, Switzerland and Great Britain has for the first time combined a photonic processor architecture, in which data is represented by laser pulses, with an in-memory architecture. The chip presented in the journal Nature can process up to two trillion combined multiplication and addition operations per second.
Software based on machine learning is capable of recognizing and translating language or analyzing images and videos to a degree that has never been seen before. But the more complex the tasks and the greater the skills, the more the computational effort increases. Although the problem can be contained by massive parallelization. A bottleneck remains, however, that a lot of data has to be transported from the memory to the arithmetic unit and back. IBM, whose research department in Zurich is also involved in this project, has therefore been experimenting with so-called “in memory computing” for years. Values are not only stored in the memory, but are also calculated directly there.
Matrix vector multiplication
A variant of this are analog accelerators for matrix-vector multiplications: this is mapped electronically via a grid-like structure. Programmable resistors form the connection between rows and columns of a grid-shaped matrix – a “crossbar array”. The electrical conductivity of these components is represented by the values of the matrix elements. Voltages that correspond to the input values are applied to the lines of the crossbars. The currents in the columns of the grid then correspond to the output values.
However, this principle has two disadvantages: On the one hand, you need easily reproducible programmable resistors – mass production of such memristors has not yet started. On the other hand, you can only multiply one matrix with an input vector in this way. In the photonic chip, however, this operation can be elegantly parallelized.
Because in this case the matrix consists of light guides, at the intersections of which a piece of phase change material with adjustable absorption is connected to one another. The researchers used a chip-based frequency comb as a light source. In contrast to the electronic version, they were able to calculate the matrix-vector multiplications for many different wavelengths in parallel.
Recognition of handwritten digits via a neural network
In order to demonstrate the efficiency of their process, the physicists constructed a “convolutional neural network” – a folding neural network for recognizing handwritten digits. With networks of this kind, a small section of the image is pushed over the entire image area and a filter is applied to this section in order to emphasize properties of the image such as edges etc. Many matrix multiplications are carried out in just one time step. The researchers came up with a processing speed of around two Teramacs (Multiply Accumulation Operations per second). In principle, however, the speed of the calculation is only limited by the bandwidth in the modulation of the laser and the measurement of the light intensities.
The researchers then let the classification, i.e. the actual recognition of the handwritten input, run on a conventional laptop. “In principle, however, it would also be possible to calculate this using photonic components,” says Johannes Feldmann from the University of Münster, main author of the study.