Building scalable networks for AI services

Having access to reliable computing power is necessary as businesses deploy and expand advanced AI applications. Neoscalers are meeting this demand by providing customized access to AI services, relieving the burden of owning and managing large amounts of expensive and complex computing and network resources. To address the infrastructure bottlenecks common with hyperscalers (e.g. Amazon Web Services, Microsoft Azure, Google Cloud), neoscalers deliver high-performance computing specifically for AI-heavy workloads.

While Ethernet has been the choice for data center networks, the high-performance computing demands of AI workloads across large, distributed GPU clusters challenges the performance, scalability and efficiency of those networks. AI applications require low-latency, high bandwidth networks to handle complex AI workloads and if the network can’t deliver the data fast enough, those expensive processors will be underutilized and less cost effective. High speed optical networks are the answer to demand for reliable, secure, scalable bandwidth.

Providers of AI compute, cloud, and edge services, data center operators hosting AI compute and interconnection, or large enterprises leveraging cloud and AI applications all need high performance, reliable optical networks to provide a reliable foundation for staying connected.
 

The Scalability Imperative in the AI Era

As enterprises ramp up their use of AI to support numerous business processes, nearly 70% of CTOs and IT leaders report that “our network infrastructure doesn’t have the capacity to embrace gen AI to its full potential”.^[1]  Traffic demands to support AI continue to rise and neoscalers customizing GPU access for businesses cannot rely on existing networks. Many are looking to AI optical networks to meet the growing demands of AI applications.

However, the introduction of an AI optical network is not a minor upgrade. Delivering purpose-built network infrastructure for AI depends on specific characteristics, and requires:

• Scalability – Highest capacity connectivity, up to 1.6Tb/s, enables faster speeds, all at less space and power
• Automation – Streaming telemetry and comprehensive support of standard open APIs enables sophisticated automation, allowing neoscalers to rapidly build and expand optical networks, activate many new nodes each day, and easily provision capacity, mirroring the rapid deployments touted by hyperscalers.
• Seamless connectivity – Capacity constraints happen everywhere from the edge of the network to the core. Neoscalers require optical solutions that can transport capacity from edge to core and provide seamless connectivity for AI in any part of the network.
• Platform innovation – Neoscalers are refining models for AI service delivery, and the network has to keep up. AI services require modern, adaptable optical networks that enable rapid introduction of new features and functions. Neoscalers need optical networks with reliable performance and scalable capacity to meet today's traffic requirements and future AI demands.

Optical networks are becoming necessary from the edge of the network to the core as AI exponentially increases network traffic. To meet this demand, neoscalers need to implement an optical network that can keep up.
 

AI Optical networking – Delivering the Speed You Need

Neoscalers are rapidly capitalizing on demand with new, high-capacity network builds. Network infrastructure is rapidly expanding from 400Gb/s to 800Gb/s and now, 1.6 Tb/s of optical capacity. Coherent optical technology provides the performance and flexibility required to transport significantly more information on the same fiber, while enabling neoscalers to transport data from the edge to the core of the network. The insatiable need for capacity will affect network expansion both within and between data centers, as well as from the edge to the core of the neoscaler network. 

Coherent optics enables greater network flexibility and programmability by supporting different rates and modulation formats providing the ability to optimize performance for any given distance, with single carrier line rates scaling to 1.6Tb/s, delivering increased data throughput at a lower cost per bit.

Bandwidth demand continues to overwhelm providers of data center interconnect services. Coherent optical solutions represent a major leap forward in communication technology, addressing the need for faster and more reliable data transmission.
 

Capacity Meets Efficiency

Ciena is already powering AI networks through its coherent optical systems. Combining breakthrough technology with its decades of experience in optical networking, Ciena’s WaveLogic coherent optics enable higher fiber capacities with unprecedented levels of automation and intelligence for real-time link monitoring and capacity tuning. In addition to the WaveLogic 5 Nano and WaveLogic 6 Nano 400Gb/s and 800Gb/s coherent pluggables, Ciena’s WaveLogic 6 Extreme (WL6e) is the first-of-its-kind proven, commercially available 1.6Tb/s optical transport system.

In addition to increased capacity, Ciena’s AI-optimized optical network infrastructure delivers:

• Spectral efficiency – Transport more capacity on deployed fiber
• Operational simplification – Deploy and manage fewer wavelengths
• Quantum safe networking – Ensure data confidentiality in the quantum era
• Sustainability and cost reduction – Double capacity with next-gen coherent using the same space and power
• Real-time visibility – Streaming telemetry data for real-time network monitoring

Lower power, less space and less cooling while increasing network capacity contributes to the business case for neoscalers as they optimize optical transport for each customer. As an industry leader in optical networking, Ciena delivers solutions and the expertise to optimize optical network design for AI while ensuring the most cost-effective outcomes.