Photonic Systems for AIDC: From Optical Fabrics to Computing and System Co-Design

Photonic Systems for AIDC: From Optical Fabrics to Computing and System Co-Design

Description: 

The explosive scaling of large foundation models is pushing AIDC infrastructure beyond the limits of current statically-interconnected electrical network architectures. To support the future AIDC’s scale-up, scale-out, and scale-across scenarios, photonics is moving toward xPU-facing optical interconnects, optical switching fabrics, and even optical computing acceleration. Emerging technologies such as high-density optical interfaces, next-generation interconnect architectures, OCS, programmable photonic fabrics, and photonic computing bring both tremendous opportunities and challenges as they step up on the central stage of AIDC.

This workshop focuses on key turning points and unresolved debates in AIDC photonic systems. It is organized in two sessions. The first session will explore how optical interconnects and switching fabrics can reshape AI infrastructure; and the second one will examine whether photonics can move toward computing and workload-aware system co-design. The workshop aims to bring together experts from photonics, AI systems, computing architecture, DC networking, and industry deployment to discuss technical feasibility, architectural trade-offs, commercialization challenges, and open ecosystems.

The key questions to address in session 1:

• GPU-facing optical interfaces and interconnect architectures: Beyond traditional paradigms, which emerging optical interface solutions and interconnect standards can best meet the power, cost, reliability, serviceability, and deployment requirements of future AI clusters?

• OCS and all-optical native AI infrastructure: Can OCS evolve from successful technical demonstrations into practical hyperscale deployment? What switching technologies (e.g., 3D MEMS, LCoS, silicon photonics), control mechanisms, and engineering solutions are needed for scale-up, scale-out, and scale-across AI infrastructures?

• Workload and traffic adaptation for optical fabrics: How should OCS and optically switched fabrics support large-model training and inference workflows? How can optical networks balance long-lived, high-bandwidth AI flows with short-packet, bursty interactive services?

The key questions to address in session 2:

• Interconnects vs. computing: Will photonics mainly serve as a high-performance data movement technology, or can optical computing approaches, such as photonic matrix–vector multiplication engines, realistically outperform or complement electronic AI accelerators?

• Software–hardware co-design: How can AI workloads, algorithms, photonic devices, optical switching fabrics, and system software be co-designed? Should AI algorithms adapt to photonic constraints, or should photonic systems evolve to fit existing computing paradigms?

• Open ecosystem, standards, and commercialization: What role should open software interfaces, orchestration frameworks, benchmarks, and standardization play in accelerating the adoption of photonic systems in AI data centers? How can the industry move from technical validation to scalable commercial deployment? Will ecosystem fragmentation or proprietary electronic alternatives (e.g., NVLink) delay the large-scale adoption of optical fabrics?

The panel aims to bring together speakers from optical networking, AI systems, photonic computing, hyperscale data centers, and standardization communities to discuss whether optical interconnects, optical switching, and optical computing will converge into a unified AI infrastructure, or evolve as competing technology paths.

Time: TBD

Venue: TBD

Organizers:

Yuyang Liu, China Telecom Chair

Xiao Lin, Fuzhou University Co-Chair

Massimo Tornatore, Politecnico di Milano Co-Chair

Wei Wang, Beijing University of Posts and Telecommunications Co-Chair

Speakers:

TBD