Photonics to satisfy the energy hunger of AI and data centers

Contents

"Exploding data volumes in all areas of life are constantly increasing the energy requirements of data centers and represent a considerable ecological and economic burden," warns the Information Technology Society (ITG) in the VDE electrical engineering and IT association in a recently published impulse paper. In particular, the currently popular generative artificial intelligence (AI) requires "enormous computing power". The ITG recommends relying on optical and photonic technologies to solve the problem. These could "play a key role in reducing the energy requirements of data centers" and at the same time meet the growing demands for speed and computing power.

"Conventional data centers still rely heavily on copper cables for data transmission," the authors of a VDE information letter write. "The transmission of large amounts of data in this way is only efficient to a limited extent and therefore energy-intensive." Copper cables exhibit signal attenuation over long distances. Amplifiers are therefore required to transmit the signal and ensure data integrity. These aids require an electrical power supply and therefore contribute to the overall energy consumption of data centers. Furthermore, the potential of alternative optical fiber technologies has not yet been sufficiently exploited.

Upgrading data channels for up to 400 GBit/s

According to the ITG, data channels that support transmission speeds of up to 400 GBit/s over a single laser wavelength in a fiber optic cable will be required in the future. To process these bandwidths more energy-efficiently, alternatives to conventional integrated digital circuits, such as integrated silicon photonic electronic circuits, would have to be addressed.

According to the paper, such photonic integrated circuits (PICs) have many advantages: The use of integrated photonics could reduce energy consumption in data centers, for example: "Light modulation can be used for communication technology with lower energy consumption." In addition, data in the high-frequency spectrum of light can be transmitted in higher bandwidths. The conversion of electrical signals into optical signals offers a number of advantages here, while at the same time increasing energy efficiency, explains the ITG. Furthermore, losses due to heat dissipation in photonic circuits are lower, which also leads to higher energy efficiency. This allows the cooling system to be significantly optimized, which in turn reduces the associated losses.

Searching with AI systems would massively increase power consumption

"Integrated photonics has already found its way into data centers, especially in high-performance computing (HPC) environments," the information states. "PICs are used to connect servers, storage devices and networking equipment. It is therefore expected that photonic interconnects for data centers will become the industry standard in the near future."

The current approach can no longer be continued, explains Damian Dudek, Managing Director of ITG. Both the training of AI systems and their operation require large amounts of energy. When a user sends a query to ChatGPT, it consumes an estimated three to nine watt-hours. If all the nine billion daily search queries were answered by AI, power consumption would increase by a factor of 30.

Nuclear power for AI doesn't have to be

By focusing on photonics, the VDE also wants to counter the efforts of large cloud service providers such as Amazon, Google and Microsoft, who increasingly hope to satisfy their AI-driven hunger for energy with nuclear power. The hyperscaler Google alone hopes to purchase energy from Small Modular Nuclear Reactors (SMRs) from the developer Kairos Energie from 2030, the authors state. The annual output should reach 500 megawatts by 2035. It is still unclear whether electricity from the reactors will be fed into the grid or whether they will be connected directly to the data centers.

In the future, the ITG expects that innovations in the field of computer hardware will also contribute to lower energy requirements and higher computing power. For example, attempts are being made to replace the conventional "Von Neumann computer architecture" used to date with approaches such as neuromorphic computing. However, these efforts alone are unlikely to be sufficient to keep pace with the power requirements of AI applications in data centers. Optical and photonic approaches for information and communication technology should therefore "make a significant contribution". In its paper, the ITG refers to considerations by engineer Thomas Ako in his Photonics Report blog from January, among other things.

(nen)