"Fit-for-Purpose or Falling Behind? Rethinking Optical Network Design for the Scale-Across Era

"Fit-for-Purpose or Falling Behind? Rethinking Optical Network Design for the Scale-Across Era

Description: 

Scale-Across — interconnecting geographically distributed GPU clusters into a unified AI computing fabric — is no longer a future scenario. Unlike Scale-Up and Scale-Out, Scale-Across imposes a fundamentally different and more demanding set of requirements on optical networks. Yet the optical networking community has not reached consensus on whether existing architectures, technologies, and design philosophies are adequate to meet this moment. This workshop brings together hyperscale AI operators, optical system and component vendors, and academic researchers to confront a direct question: is the optical network being designed for Scale-Across, or is it falling behind?

The following topics are proposed as the scope of this workshop.

Topic 1｜What Does Scale-Across Really Demand from Optical Networks?

Scale-Across spans a wide distance continuum — from sub-kilometer intra-campus links, to metropolitan-scale connections of 10–80km, and wide-area interconnects exceeding 100km — and across this entire range, the demand for high bandwidth is certain, but many other dimensions remain open for debate: how critical is deterministic latency for AI training workloads, and is jitter truly a problem that needs to be solved at the optical network layer, or can it be absorbed by upper-layer protocols and AI frameworks? How should client-side port rates evolve — 800G or 1.6T, single-port or multi-port aggregation, and on what timeline? RDMA is the dominant communication protocol for AI clusters — what is its relationship with the optical transport network, does the optical layer need to be aware of RDMA traffic characteristics, or can the two remain completely decoupled?

Topic 2｜Multi-Rail and Petabit Capacity: How Should Optical Network Architecture Evolve?

The push toward petabit-scale capacity is a certain direction for Scale-Across optical networks, but the path to get there is filled with open questions: is Multi-Rail an indispensable architectural choice, or an unproven concept yet to be validated at scale — what problem does it actually solve, and what additional cost and complexity does it introduce? As the electrical layer continues to evolve with ever-higher forwarding capacity, what will the next-generation transponder look like, and what level of chassis integration density is achievable? In point-to-point Scale-Across scenarios, is the wavelength flexibility provided by WSS still necessary, or is a fixed-grid architecture sufficient? Can C+L band expansion meet future capacity demands, or does reaching true petabit-scale require the introduction of the S band or even spatial division multiplexing?

Topic 3｜Protection for Scale-Across: Is What We Have Enough, or Does It Need to Be Redefined?

Protection and resilience are unavoidable topics for Scale-Across optical networks, but the answers are far from settled: what does traditional 50ms optical-layer protection switching actually mean for an AI training job — does it truly constitute a problem, or do upper-layer systems have the ability to tolerate it? Does Scale-Across require entirely new protection schemes beyond existing 1+1 and shared mesh approaches, or is optimization of current mechanisms sufficient? Should the optical transport layer and the data communications layer jointly coordinate protection decisions, or is it more robust to keep them operating independently? At what granularity should protection be defined — fiber, wavelength, cable, or flow level? Is lossless switchover technically feasible, and if so, can it deliver meaningful business value for AI workloads — and at what cost?

Topic 4｜Autonomous Management: Does a Scale-Across Optical Network Really Need It?

Autonomous management is frequently cited as essential, but its necessity and feasibility remain open questions: for an optical network purpose-built for Scale-Across, is autonomous management a nice-to-have or an operational prerequisite? If building a greenfield network from scratch with no legacy constraints, does that make autonomous management easier to achieve — and where do the real barriers lie, in technology or in organizational process? What northbound interface capabilities does the equipment need to expose, and are existing interface standards already sufficient? Across the spectrum from alarm reporting and assisted recommendation to fully closed-loop autonomy, what level of automation is trustworthy and deployable today? What telemetry sampling granularity and real-time performance are required, and can current optical network equipment support this natively, or does it demand additional infrastructure investment?

Time: TBD

Venue: TBD

Organizers:

TBD

Speakers:

TBD