A Russian photonic computer could become a partial specialized alternative to GPUs for specific AI operations. This conclusion was reached by participants of a round table, TASS was informed by the press service of Element Group. This refers to inference, signal processing, and optimization.
This requires a hybrid circuit: a photonic matrix core, electronic control and memory, a light source, packaging, testing, PDK, and a software stack.
Yury Sementsov, Vice President for Strategic Development at Element, noted that the photonic core does not require 3–5 nm topologies — the main elements of silicon photonics can be implemented on mature 90–45 nm processes, and electronics can be assembled separately. This is what makes photonics a realistic direction for the domestic production base.
In the US, photonic interconnects for AI clusters are being implemented first, with computing modules entering the market in the 2028–2029 timeframe. Developers claim an energy efficiency gain of more than an order of magnitude, but for now, these are mostly their own estimates. Issues of memory, signal converters, and the accuracy of analog computations remain unresolved.
In Russia, the Moscow Photonics Center operates — the first industrial site for photonic integrated circuits. Planar photonics technologies on silicon nitride are being developed at the Institute of Nanotechnology of Microelectronics of the Russian Academy of Sciences, and the Laser Center with Alferov Academic University has created a serial lithographer for maskless laser lithography. Sber is studying which neural network operations to assign to photonics and has already shown the first prototypes of optical computers.
By 2027, requirements for the project need to be formulated, and by 2028–2029, the module should be brought to pilot operation and demonstrate a task with a measurable advantage over GPUs.