NVIDIA Unveils BlueField-4 to Power Scale-In AI Networks

NVIDIA has launched its BlueField-4 data processing units (DPUs) as the foundation for "Scale-In" networking, a new architectural pillar designed to secure and manage traffic in agentic AI factories. The BlueField-4 DPU features a 64-core NVIDIA Grace CPU that delivers six times more compute power than its predecessor, alongside LPDDR5X memory and a PCIe Gen6 host connection. Compared to the BlueField-3, the new chip offers four times more memory bandwidth and double the network bandwidth, supporting throughput speeds of up to 800 Gb/s.

This hardware integrates with NVIDIA DOCA software and Spectrum-X Ethernet to handle north-south data movement without taxing host processors. When deployed in the NVIDIA Vera Rubin NVL72 system, the architecture coordinates 7.2 Tb/s of aggregate interface bandwidth. This setup combines an 800 Gb/s north-south BlueField-4 path with four 1.6 Tb/s east-west paths per compute tray, where ConnectX-9 SuperNICs manage tenant workload traffic. For storage operations, the co-designed Spectrum-X and BlueField-4 pipeline delivers up to 1.45x more storage throughput than standard, off-the-shelf Ethernet.

For system administrators and developers, this release shifts critical infrastructure management entirely outside the tenant host. Using the NVIDIA BlueField Astra control point, operators can provision isolated virtual private clouds (VPCs) and enforce security policies across both north-south and east-west fabrics. Software tools like DOCA Host-Based Networking (HBN) and OVS-DOCA manage routing and isolation, while DOCA Argus handles runtime threat detection and DOCA Vault secures file access.

Practitioners can also automate deployment using the Kubernetes-native DOCA Platform Framework (DPF) to discover, provision, and update DPUs across the fleet. Meanwhile, DOCA Telemetry gathers real-time performance data independent of the host operating system. This separation ensures that security policies and telemetry collection remain tamper-proof, allowing operators to resolve bottlenecks and maintain high effective bandwidth without degrading GPU performance.