F5 BIG-IP NEXT for Kubernetes delivers 3x performance gains in GPU cluster efficiency through intelligent load balancing.

We toured a lab featuring F5's BIG-IP application delivery and security stack running on NVIDIA BlueField DPUs. While the solution can also run on host CPUs, that approach consumes valuable processing cores. F5 BIG-IP Next for Kubernetes (BNK) sits between clients and GPU infrastructure to handle LLM routing, token governance, and Layer 4 to Layer 7 security at the network edge. We observed its performance impact firsthand, which is meaningful—finding ways to increase GPU performance translates directly into more throughput from the same power and GPU footprint, both critical constraints today.

We traveled to California to visit the lab and film on-site. This was only possible through F5's sponsorship, as the facility normally restricts outside access. The lab itself was exceptionally clean and well-organized.

F5 manufactures both hardware appliances and software solutions. The BNK offering brings much of what traditionally ran on dedicated appliances onto NVIDIA BlueField-3 DPUs, or alternatively onto host CPUs.

Documenting networking infrastructure in AI racks matters because many underestimate its complexity. GPU clusters require multiple distinct networks: GPU-to-GPU traffic, GPU-to-storage communication, user and application networks, control planes for nodes and host operating systems, and networks for switches, cooling infrastructure, PDUs, and serial console servers. Some operate at lower speeds; others run at 400Gbps or 800Gbps per port, increasingly standard for high-speed fabrics.

Our test cluster required several GPU servers to simulate multiple concurrent users. We deployed NVIDIA H100 "Hopper" generation GPUs in Supermicro servers, along with additional systems for Kubernetes hosts and cluster storage.

Each Supermicro GPU server hosted two NVIDIA BlueField-3 DPUs, while the top 2U systems each carried eight DPUs. Even the DPUs themselves maintain separate management interfaces, adding another layer to the infrastructure.

The sheer volume and diversity of networking in modern AI clusters—spanning different traffic types, multiple user profiles, and various applications—requires sophisticated solutions. We equipped our lab with Keysight IxNetwork and CyPerf load generators, each capable of generating over 1.6 terabits per second of traffic. This visit underscored how critical it is to manage application delivery, user access, and security across this complexity.

We now proceed to test the F5 BIG-IP Next for Kubernetes solution.