Though Juniper Networks is in the process of being acquired by HPE, the company's product innovation is still going strong.

Juniper Networks is growing its portfolio today with a series of updates that the company claims will enable AI-driven networking. Central to the platform is an expanded version of Juniper's Marvis Virtual Network Assistant (VNA). Marvis now features digital twin capabilities through Marvis Minis, which uses artificial intelligence (AI) to simulate user connections and proactively validate network configurations without real traffic. Additionally, Marvis VNA now supports data centers, providing cross-vendor hardware insights and a conversational interface via AI.

Complementing Marvis is new hardware built for AI workloads. Juniper is also rolling out new PTX routers with Express 5 silicon and new QFX switches with Broadcom Tomahawk 5 silicon to enable high-density 800G capacity for AI infrastructure. Juniper is also expanding its Apstra data center solution for streamlined deployment and assurance of AI training and inference clusters.

“This launch is a commitment from Juniper that we’ll be embedding AI into our platform from the beginning,” Jeff Aaron, VP of Enterprise product marketing at Juniper Networks, told SDxCentral. “More specifically, AI-native means we pull the right data for AIOps across all areas of the network, have a secure cloud infrastructure designed for AI, and have the only VNA driven by 7 years of AI learning.”

Juniper's AI networking path is a bit Mist(y)

Juniper has been talking about AI networking long before the current hype cycle surrounding generative AI (genAI).

Back in 2019, Juniper acquired Mist Systems for $405 million, bringing along with it AI capabilities and the Marvis platform. Juniper has been steadily improving Marvis in the years since with intelligence updates.

Aaron noted that before this week, the Marvis VNA was only present in campus and branch networking technologies from Juniper. Now, it’s in the data center, too, for end-to-end visibility and assurance. Aaron explained that with the new Marvis for data centers, organizations can get insight and control into key actions pertaining to data center connectivity and performance. The system also provides a genAI conversational interface to query tech documentation and other knowledge base items.

Marvis Minis bring new insight to AIOps

The new Marvis Minis technology brings a digital twin approach to Juniper's AIOps strategy.

Aaron commented that in the past, Marvis could detect and fix problems rapidly after users encountered an issue, without waiting for a user to report an issue. In some instances, the users may have not known the problem existed, but it still was primarily in reaction to a user issue such as inability to connect. This then could be translated into proactive and predictive actions for the future, but again, the original source was a user-triggered event.

“Now, we can become even more proactive by detecting and fixing issues without users ever having to be present at all,” he said.

By simulating users and IoT devices, Aaron explained that Juniper can validate configuration changes immediately, find and fix problems when no one is on the network, or before users on the network encounter an issue.

“This provides even more insight into the network, assures even better network performance and ultimately results in better end-user experiences,” Aaron said.

How Juniper Apstra is getting a boost for AI

Another key element of Juniper's AI strategy was also gained via acquisition, again years before the current AI hype cycle.
In 2020, Juniper acquired intent-based networking vendor Apstra and has been steadily improving the technology in the years since to become a fabric management and automation platform.

Aaron said that for AI/machine learning (ML) workloads, Juniper Apstra has been expanded to provide faster and more efficient processing of AI/ML traffic over Ethernet, including congestion management, load balancing and flow control.

AI-powered networking needs hardware

AI puts new challenges on compute, network, and storage solutions with large models that run in parallel across many GPUs for training. These models require fast job completion time with minimal delays for the last GPU to finish its calculations — that is, low tail latency. According to Aaron, the Juniper AI data center networking platform continues to tackle these challenges with software and hardware.

Among the new hardware elements announced today is the Juniper Networks QFX5240 800GbE Switch, which is a next-generation, fixed-configuration platform designed for spine, leaf and border Switch roles. Aaron said that the Switch provides flexible, cost-effective, high-density 800GbE, 400GbE, 100GbE and 50GbE interfaces for intra-IP fabric connectivity as well as higher density 200/400GbE network interface connectivity for AI/ML use cases.

Juniper is also continuing to boost its Ethernet capabilities to support AI workloads. To that end, the QFX uses Remote Direct Memory Access (RDMA) over Converged Ethernet v2 (ROCEv2) for transport at the network layer.

“QFX5240 supports ROCEv2 along with congestion management features like PFC, explicit congestion notification (ECN), and data center quantized congestion notification (DCQCN),” Aaron said.