AI data center networking

Organizations need well-designed data center networks to tackle the challenging demands of AI workloads. These networks must extend seamless, reliable connectivity across the AI infrastructure and deliver the best possible performance for every AI training and inference task.

The back-end network is used to interconnect the high-value graphics processing unit (GPU) resources required for high-computation tasks such as AI training, AI inference or other HPC workloads. It delivers lossless, low-latency and high-performance connectivity for AI training compute and dedicated storage resources. 

The front-end network supports connectivity for AI workloads, general-purpose workloads (non-AI compute) and the management of AI workloads. In the context of AI inference, the front-end delivers fast response times and proximity to the end user. 

In both cases, it’s essential for organizations to control—and, ideally, reduce—their cost and power use as they deploy networks to meet these new demands. 

More and more organizations are choosing Ethernet technologies to build back-end networks for AI workloads. Ethernet already dominates front-end network designs. The work of the Ultra Ethernet Consortium (UEC) will bring enhancements that make Ethernet the best choice for AI network infrastructures. 

UEC is working to deliver an open, interoperable, high-performance, full-communications-stack architecture based on Ethernet to meet the growing network demands of AI and HPC at scale.

UEC members, including Nokia, aim to leverage the ubiquity, performance curve and cost benefits of Ethernet to evolve the legacy RDMA over Converged Ethernet (RoCE) protocol with Ultra Ethernet Transport (UET). This modern transport protocol is designed to enhance network performance to meet the requirements of AI and HPC applications while preserving the advantages of the Ethernet/IP ecosystem.

The Nokia Data Center Fabric solution provides the reliability, simplicity and flexibility you need to build and deploy network infrastructures that can meet the requirements of current and future AI workloads.
 

Modular and fixed configuration platforms

Our solution includes a comprehensive portfolio of modular and fixed-configuration hardware platforms for implementing high-performance leaf–spine designs. You can use our platforms to build high-capacity, low-latency and lossless backend networks that can efficiently handle demanding AI training workflows. We also offer platforms in many different form factors to support frontend network designs that will interconnect your AI inference compute, non-AI compute and shared storage resources.
 

A powerful and proven network operating system

Our data center hardware platforms run on the Nokia SR Linux network operating system (NOS). SR Linux opens the NOS infrastructure with a unique architecture built around model-driven management and modern interfaces. It is designed for reliability and quality, ready for automation at scale, and easy to customize and extend. 

SR Linux provides congestion management and traffic prioritization capabilities that let you deliver lossless Ethernet networking. It also supports high-performance AI infrastructures with superior telemetry, manageability, ease of automation and resiliency features.
 

Fabric management and automation toolkit

Nokia Event-Driven Automation (EDA) is a modern data center network automation platform that combines speed, reliability and simplicity. It makes network automation more trustable and easier to use wherever you need it, from small edge clouds to the largest data centers.

With EDA, you can automate the entire data center network lifecycle from initial design to deployment to daily operations. This allows you to ensure reliable network operations, simplify network management and adapt to evolving demands.

EDA offers a range of flexible AIOps capabilities through its natural language interface. Operators simply use this interface to ask an operational question in the areas of troubleshooting, root cause analysis, remediation support, and more, and through an Agentic AI process, the system responds.
 

Data center interconnectivity

Some companies may implement their own training infrastructures. Others may prefer to use AI platforms and associated services from large cloud providers, an approach known as GPU as a service (GPUaaS) or public AI.

Inferencing models that need low-latency end-user access are typically run in enterprise edge locations in private AI infrastructures. Batch-mode and global-scale inferencing models are better suited for implementation in public AI frameworks.

We offer a comprehensive solution that makes it easy to interconnect AI infrastructures between data centers and across the WAN. The Nokia Optical Data Center Interconnect (DCI) and Nokia Data Center Gateway (DCGW) solutions let you provide reliable, high-performance interconnection across AI infrastructure domains. They can help you meet evolving distributed AI connectivity requirements so you can serve AI models at the edge of the network, closer to end users.