Juniper multivendor lab designed to maximise AI performance

Author: Simon Rowley

Juniper Networks, a provider of secure, AI-native networking, today announced the first and most comprehensive multivendor lab for validating end-to-end automated AI data centre solutions and automated operations with switching, routing, storage and compute solutions from leading vendors, as well as new Juniper Validated Designs (JVDs) that accelerate the time-to-value in deploying AI clusters.

In addition, Juniper is releasing new key software enhancements that optimise the performance and management of AI workloads over Ethernet. Through these Operations for AI – Ops4AI – initiatives, Juniper is collaborating closely with a broad range of infrastructure ecosystem partners to enable the best AI workload performance via the most flexible and easiest-to-manage data centre infrastructures.

As a key element of Juniper’s AI-native networking platform, the existing networking for AI solution consists of a spine-leaf data centre architecture with a foundation of AI-optimised 400G and 800G QFX series switches and PTX series routers. The solution is secured via high performance firewalls with industry-leading effectiveness, and managed via Juniper Apstra data centre assurance software and the Marvis Virtual Network Assistant (VNA).

Juniper Apstra and Marvis provide key Ops4AI capabilities, such as intent-based networking, multivendor switch management, application / flow / workload awareness, AIOps proactive actions and a GenAI conversational interface. With Juniper’s full networking for AI solution, customers and partners can lower AI training Job Completion Times (JCTs), reduce latency during inferencing and increase GPU utilisation while decreasing deployment times by up to 85% and reducing operations costs by up to 90% in some instances.

To simplify AI clusters and maximise network performance even further, Juniper has added new Ops4AI software enhancements that together offer unique value for customers. The enhancements being announced today include:
• Fabric autotuning for AI: Telemetry from routers and switches are used to automatically calculate and configure optimal parameter settings for congestion control in the fabric using closed-loop automation capability in Juniper Apstra to deliver peak AI workload performance.
• Global load-balancing: An end-to-end view of congestion hotspots in the network (i.e. local and downstream switches) is used to load-balance AI traffic in real-time, delivering lower latency, better network utilisation and reduced JCTs.
• End-to-end visibility from network to SmartNICs: Provides an end-to-end holistic view of the network, including SmartNICs from NVIDIA (BlueField and ConnectX), and others.

Industry’s first multivendor Ops4AI lab to collaborate with ecosystem and validate operations

Openness and collaboration are core to Juniper’s networking mission as they are the only way to move AI data centres from their current early adopter stage to effective mass market deployments. End-to-end operations for multivendor AI data centre infrastructure has been difficult, leading to vertically integrated AI data centre solutions that are vendor-locked and lead-time challenged.

As a result, Juniper has launched the industry’s first Ops4AI Lab with participation from Juniper’s partner ecosystem including Broadcom, Intel, NVIDIA, WEKA and other industry leaders. The Ops4AI Lab, located at Juniper’s Sunnyvale, California corporate headquarters, is open for all qualified customers and partners who want to test their own AI workloads using the most advanced GPU compute, storage technologies, Ethernet-based networking fabrics and automated operations. Ops4AI Lab testing using validated Ethernet fabrics delivers comparable performance to InfiniBand-based AI infrastructure.

Users requesting a slot in the Juniper Ops4AI lab should contact their local Juniper Networks sales team.

Juniper Validated Designs to provide assurance

Juniper Validated Designs are detailed implementation documents that give new customers confidence that the solution and topology they have chosen is well characterised, well tested and repeatable, resulting in faster time to successful deployment. All JVDs are proven integrated solutions, tested in best practice designs based on specific platforms and software versions.

Juniper has released the first pre-validated blueprint specifically for AI data centres, built on NVIDIA A100 and H100 compute, storage from Juniper’s ecosystem partners, and Juniper’s portfolio of data centre leaf and spine switches. This new Ops4AI JVD complements Juniper’s existing JVDs for automated, secure data centres which include QFX and PTX spines, QFX leaf switching, data centre automation, and Juniper’s SRX and vSRX/cSRX solutions for data centre security.

Praveen Jain, SVP and GM of AI and Data Centre at Juniper Networks, comments, “Best-of-breed always wins out, and the same will be true for compute, storage, networking and operations in AI data centres. Juniper has made a significant investment in the Ops4AI lab, JVDs, and a new promotional programme to enable our customers and partners to have maximum choice, flexibility and stability in how they build a complete GenAI solution. There has never been a better time to build high-performance, low-latency, multivendor AI data centre solutions that are simple, fast and economical to deploy and operate.”

For more from Juniper, click here.