Sber presented Russia's first universal optical computer for artificial intelligence tasks at the SPIEF. Its key feature is that it uses light for calculations instead of conventional electronics. The device is based on a photonic integrated circuit, the architecture of which was developed by the company's researchers.
Such chips are needed due to the growing load from neural networks. The most energy-intensive part of AI's work is matrix computations, which are used in training and running large models. In an optical computer, these operations are performed almost instantly and with less heat generation.
According to Andrey Belevtsev, Senior Vice President of Sber, the first prototype is already capable of performing more than 1 billion matrix multiplication operations per second. The company sees the possibility of increasing the frequency of optical operations to 10 GHz and higher, and the energy consumption of the optical core is already more than 30% lower than electronic counterparts.
The development could be important for data centers, where energy costs due to AI are growing ever faster. Sber considers photonics one way to make computing infrastructure more efficient and scalable. In addition, according to the company's estimates, the production of optical chips is easier to localize than electronic ones.
If the technology develops, businesses will be able to train and use large AI models more cheaply. In essence, Sber has shown not just an experimental chip, but an attempt to find a new path for computing in an era when neural networks require more and more power and less and less unnecessary energy consumption.