Optical Interconnect Technologies for AI Networks

Optical Interconnect Technologies for AI Networks

🎙 Dr. Tzu Hao Chow 👥 5K 📅 January 30, 2026 ⏱ 24 min 👁 97 📄 expert opinion 🧭 2026-08-16
Available in: English (current) Français

Keywords

optical interconnectAI scale-upco-packaged opticsVCSELsilicon photonics

Summary

Dr. Tzu Hao Chow from Broadcom presents on optical interconnect technologies for AI networks at SEMICON Taiwan 2025. He begins by outlining the current landscape: optical interconnects are well-established in front-end and scale-out networks, but scale-up networks (GPU-to-GPU) have traditionally relied on copper. As data rates increase, copper reaches its limits in reach and power efficiency, motivating the adoption of optics. He introduces two key solutions: VCSEL-based multi-mode optics (MO) and silicon photonics-based co-packaged optics (CPO). For VCSEL MO, he highlights its low power consumption (1 pJ/bit), cost competitiveness, and reliability (sub-0.1 FIT). For silicon photonics CPO, he emphasizes its high bandwidth density (up to 51.2 Tbps) and longer reach (up to 2 km), but notes higher power. He discusses reliability testing, link stability, and the importance of cost, power, and bandwidth density in comparing against copper. He concludes that both technologies will play roles, with CPO promising for scale-up networks, and emphasizes the need for proven reliability and link stability for adoption.

163 words

Critical Evaluation

Value of the Information & Strength of the Argument

The presentation provides valuable insights into the technical and economic considerations for adopting optical interconnects in AI scale-up networks. The speaker presents a clear comparison between copper and optical solutions, backed by specific metrics such as power efficiency (pJ/bit), reach, and bandwidth density. He argues that optics can overcome copper’s limitations in reach and power, enabling larger cluster sizes. The argumentation is solid, with data points from internal testing and industry experience. However, some claims are forward-looking (e.g., 2028-2029 projections) and not yet proven in the field.

Scientific Rigor, Source Quality, Title Accuracy

The presentation is based on the speaker’s expertise and Broadcom’s internal data. No external sources are cited. The title accurately reflects the content. The speaker mentions reliability data from ‘our site and from our peers’ but does not provide specific references. The talk is an expert opinion rather than a peer-reviewed study, so the scientific rigor is moderate.

160 words

Title / Content Match

The title accurately reflects the content, which focuses on optical interconnect technologies for AI networks.

Quality & Reliability

7/10

Presentation by an industry expert from Broadcom, providing technical details and data on optical interconnect technologies. Claims are based on internal testing and industry experience, but not peer-reviewed. Some projections are forward-looking.

Key Moments

Contribution & Novelties

The presentation provides an industry perspective on the adoption of optical interconnects for AI scale-up networks, comparing VCSEL-based MO and silicon photonics CPO. It highlights specific metrics such as power efficiency (pJ/bit), reach, and bandwidth density, and discusses reliability and link stability as key factors. The talk offers insights into Broadcom’s roadmap and testing results.

Pour aller plus loin :

88 words

Radar Profile

The radar profile shows high scores in technical level and information quality, indicating a technically detailed presentation. The lower score in information quantity suggests a focused scope. Overall, the presentation is well-balanced, with strong technical content but limited external sourcing.

Reliability 7/10