QFX5240 Switches Datasheet

The QFX5240 Switches meet the advanced AI data center networking requirements of large-scale clusters. QFX5240 switches work with the automation in Juniper Apstra to assure daily operation in AI and ML workload training and access.

QFX5240 Switches:

— deliver high-density 800GbE ports on a fixed form factor with software to provide advanced network services tuned to the specific needs of AI/ML workloads

— are a foundation of AI networks, ensuring fast job completion time (JCT) to speed training through high GPU utilization

— help teams managing AI/ML environments realize improved economics

Continuous evolution of AI/ML technology along with new applications are driving the next major shift in bandwidth requirement within the data center fabric. Juniper Networks^® QFX5240 800GbE switch (64 port 800GbE) is a next-generation, fixed-configuration platform designed for spine, leaf, and border switch roles. 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 NIC connectivity for AI/ML use cases. It’s 51.2Tbps unidirectional throughput meets the bandwidth requirement of AI/ML workloads and storage systems with latency in the range of 700-750ns. Remote Direct Memory Access (RDMA) is the de-facto data transfer technology used in AI/ML workloads, and it uses Remote Direct Memory Access 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).

The QFX5240 helps reduce the number of network nodes deployed—decreasing the total power consumption of the data center fabric and improving the carbon footprint of the data center.  These improvements are possible by having different breakout options like 128x400GbE and 256x200GbE/100GbE.

Artificial intelligence 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 (JCT) with minimal delays for the last GPU to finish its calculations, that is, low tail latency. Architects optimize the cluster performance through rail-optimized design (Read this Juniper White Paper for more information about AI/ML cluster design). As model sizes and datasets continue to grow, designs must accommodate more GPUs in the cluster, requiring that the network seamlessly scale, without compromising performance, or introducing communication bottlenecks.

The QFX5240 meets the needs of these large-scale AI networks. The switch provides:

• 64 ports of 800GbE on a 2 U switch to reduce costs on both space and total power utilization
• Choice of connectivity with both OSFP and QSFP-DD variants of 800GbE for leaf-spine connectivity
• Availability of 2x400GbE per 800GbE port for increased connectivity to the GPUs
• Advanced telemetry capabilities to support ECN/PFC counters
• Fine-grained, load-balancing capability to handle reduced flow entropy
• Automation of rail-optimized design through Apstra

Automation tools, such as Apstra, ensure the reliable set up of expansive networks with ongoing verification of the deployment along with monitoring of operations. Apstra intent-based networking delivers full Day 0 through Day 2+ capabilities for IP/EVPN fabrics with closed-loop assurance in the data center. Apstra provides a broad set of operational capabilities, with multiple built-in intent-based analytics probes, flow visibility, and analysis to ensure that the AI network is running as designed. Apstra provides a simple UI workflow to create custom intent-based analytics to capture, enrich, and visualize data from the AI network.

The QFX5240 supports Junos^® telemetry interface, a modern telemetry streaming tool that provides performance monitoring in complex, dynamic data centers. Streaming data to a performance management system lets network administrators measure trends in link and node utilization and troubleshoot issues such as network congestion in real time.

Junos telemetry interface provides:

• Application visibility and performance management by provisioning sensors to collect and stream data and analyze the application and workload flow path through the network
• Capacity planning and optimization by proactively detecting hotspots and monitoring latency and microbursts
• Troubleshooting and root cause analysis via high-frequency monitoring and correlating overlay and underlay networks

Additionally, the Junos Evolved operating system supports a robust API set to support automation through Terraform, Ansible, zero-touch provisioning (ZTP), operations and event scripts, automatic rollback, and Python scripts.