Professionals in AI, networking, and optical technology interested in the future of high-performance computing infrastructure.
TL;DR
This panel discusses "scale out" networks, crucial for AI connectivity, focusing on bandwidth, power, and cost. Experts explore how to achieve massive scale by integrating "scale up" networks, leveraging technologies like EON and advanced optical solutions to overcome limitations of traditional approaches.
Key Takeaways
Scale-out networks are crucial for AI connectivity, characterized by a broad, interoperable ecosystem focused on bandwidth, power, and cost.
Traditional scale-out networks using multi-layer Ethernet switches struggle with high radix requirements for large AI clusters, leading to excessive power consumption.
Leveraging scale-up networks, which offer high bandwidth and low latency interconnects, can be integrated into scale-out strategies to create larger, more efficient domains.
The EON standard allows scale-up and scale-out systems to use the same messaging, enabling a 'pod' approach where a single Ethernet switch layer can connect millions of GPUs.
Optical technologies are essential for scale-up networks beyond copper's reach, with ongoing advancements in data rates and power efficiency being critical.
Matching switch capabilities with optical solutions is vital; for instance, a 100 Tb switch requires optics that don't exceed 200 Gb per fiber.
Bi-directional (BiDi) optics offer an easy win by reducing fiber count without impacting radix, a proven technology for increasing efficiency.
While 'fast and wide' optics are theoretically appealing, they are currently unrealistic due to limitations in available switch silicon and complexity.
Questions & Answers
What is the main focus of the Scale Out session at OFC 2026?
The Scale Out session focuses on the economic center of AI connectivity, addressing the epic demand across the supply chain and the key factors of bandwidth, power, and cost.
What characterizes Scale Out networks?
Scale Out networks are characterized by a broad, standardized, and interoperable ecosystem, with a relentless focus on bandwidth, power, and cost.
What are the emerging optical technologies for Scale Out networks?
Emerging technologies include CPO (Common Power Option) and NPO (New Power Option), moving beyond traditional pluggables to deliver technology at an unprecedented scale.
How can Scale Up networks be leveraged for Scale Out systems?
Scale Up networks can be treated as part of the Scale Out network, allowing a single Ethernet switch layer to connect a million GPUs/XPUs within a pod structure.
What are the challenges with traditional optical technologies for Scale Out?
Traditional DR-based optics face limitations in data rate beyond 400 gig, and power efficiency is impacted by copper traces and the optics themselves, leading to high IO power.
What is BiDi and its benefit for Scale Out networks?
BiDi (Bi-Directional) technology uses the same fiber for both directions of transmission, reducing fiber count without impacting radix, making it an easy win for Scale Out.
Key Terms
Scale Out — Refers to networks designed for massive AI connectivity, focusing on bandwidth, power, and cost efficiency across a broad ecosystem.
CPO (Common Power Option) — An emerging optical technology for network connectivity, moving towards integrated power solutions beyond traditional pluggable modules.
NPO (New Power Option) — An emerging optical technology for network connectivity, similar to CPO, focusing on integrated power solutions.
BiDi (Bi-Directional) — A technology that allows data transmission in both directions over a single fiber, reducing the number of fibers needed.
Download or copy the punctuated YouTube transcript (Markdown)