How F5 BIG-IP Next on NVIDIA BlueField-3 DPUs Boosts AI Cluster GPU Performance

GPUs remain one of the most expensive components deployed in AI infrastructure, so squeezing more performance out of existing hardware is a top priority. This visit to the F5 lab in Sunnyvale, California (sponsored by F5) explores how running F5 BIG-IP Next for Kubernetes directly on NVIDIA BlueField-3 data processing units can significantly improve AI application throughput while adding security capabilities, with benchmark comparisons against an Envoy AI gateway.

■ Lab Tour and AI Cluster Architecture
- Common AI cluster components: management NICs, high-speed NICs, Dell and Supermicro servers, storage nodes
- Multiple NVIDIA BlueField-3 DPUs and NVIDIA H100 GPU servers in a Supermicro AMD EPYC 2U configuration

■ F5 BIG-IP Next for Kubernetes with BNK Reference Architecture
- AI load balancing with EPP, LM routing coordination, and token management
- Zero-trust security, multi-tenant isolation, and MCP/AI agent support
- GPU-aware endpoint selection using metrics like load, request size, and KV cache locality

■ Performance Benchmarking Methodology
- VLLM cluster with 8 GPUs, Qwen3-32B-FP8, tested with NVIDIA performance tooling
- Four traffic patterns: heavy context reuse, multi-turn chat, mixed traffic, and unique requests
- F5 versus Envoy gateway comparison under varying concurrency levels

■ DPU Hardware Deep Dive
- BlueField-3 specs: 16 ARMv8.2 cores, on-package DDR5, programmable accelerators, up to 400 Gbps networking

■ Practical Value for AI Operators
- Realizing higher utilization from existing GPU fleets without buying additional GPUs

Ideal for AI infrastructure engineers, platform teams, and IT decision makers evaluating gateways, DPUs, and GPU efficiency strategies. After watching, you will understand where gateway-level optimizations deliver the most impact and which workloads are worth retesting in your own environment.