Demand for computing power in AI servers and data centers is surging, with Silicon Optics and CPO now in practical use.

2025/05/09

Commercial Times
May 2, 2025
Author Commercial Times Li Juanping

The annual global optical communications event "OFC 2025" was held in San Diego, USA a few days ago. Driven by the generative AI craze, high-bandwidth, high-integration, and low-power transmission technology became the focus of the exhibition.

Industry insiders observed that after the introduction of multimodal models into generative AI, the demand for computing power in AI servers and data centers increased dramatically, pushing high-speed optical communications, CPO (co-packaged optics), silicon photonics and optical I/O technologies into the actual application stage.

Generative AI has become an indispensable infrastructure, especially driven by applications such as ChatGPT and Gemini. The introduction of multimodal large models has increased the demand for computing power by multiples.

In order to meet the needs of training and inference, data center operators are accelerating the adoption of higher-speed connection architectures such as 1.6T or even 3.2T. In response to higher bandwidth and distance requirements, optical communication systems are moving towards multiple solutions such as active cables (AEC), active optical fibers (AOC), silicon photonic transceivers and CPO.

In terms of connecting wires, short-distance applications (within 5 meters) are still dominated by AEC. For example, Marvell and Amphenol jointly demonstrated a 5nm DSP solution that supports 1.6Tbps transmission speed, which will be upgraded to 3nm in the future to reduce energy consumption by 20%.

Within 30 meters, AOC is mostly used because of its simple construction, low cost and energy saving, and it can already support 1.6T. As for long-distance applications over 30 meters, silicon photonic transceivers are the mainstream. Currently, 800G demand is strong, and 1.6T components have also been mass-produced.

One of the biggest highlights of this exhibition is CPO, which directly integrates optical modules with Switch ASIC or xPU to greatly improve connection density and energy efficiency.

According to legal analysts, CPO is an important technology to cope with the decentralization of AI computing architecture and the increase in bandwidth density. Nvidia, Broadcom and Marvell all exhibited relevant CPO designs.

Optical leader Coherent also displayed complete peripherals, such as high-temperature resistant connectors, polarization-maintaining optical fibers, microlenses and external laser modules (ELS). Taiwan-based SAN JOSE and Japanese company Sumitomo Electric Industries also entered mass production preparations at the same time.

Intel, Ayar Labs, Lightmatter and other major manufacturers and startups also demonstrated solutions for directly integrating photon engines on xPU chips.

Optical communication industry players pointed out that with the rapid upgrade of AI server and switch architectures, the internal transmission rate of data centers will move from 800G to 1.6T or even 3.2T in the next two years, and the demand for optical communication modules and key components will increase significantly.

【Disclaimer】
The content of this article only represents the author's personal opinion and has nothing to do with Creating.
The content, textual description and originality have not been verified by this site. This site does not make any guarantee or commitment to this article and all or part of its content, authenticity, completeness, and timeliness. It is for readers' reference only. Please verify the relevant content yourself.

Creating Nano Technologies, Inc.
59 Alley 21 Lane 279, Chung Cheng Road, Yung Kang City, Tainan, TAIWAN
TEL:886-6-2323927 FAX:886-6-2013306 URL: http://www.creating-nanotech.com

BACK TO LIST