NVIDIA Adds BlueField-4 Scale-In for 800 Gb/s AI-Factory Infrastructure Offload

NVIDIA says its new Scale-In layer is designed to prevent infrastructure around AI compute from becoming a bottleneck. It combines BlueField-4 DPUs, DOCA software and Spectrum-X Ethernet to handle access, security, storage and telemetry outside host CPUs. BlueField-4 supports up to 800 Gb/s, according to the company.

Scale-In is NVIDIA’s fifth AI networking pillar. It sits alongside Scale-Up, which links GPUs; Scale-Out, which connects servers; Scale-Across, which connects distributed AI factories; and Context Memory, which NVIDIA positions for shared KV-cache storage and reusable inference state.

The distinction is architectural. North-south networks carry traffic between an AI factory and its users, applications, enterprise data, services and external storage. NVIDIA’s argument is that those services need their own processing domain as AI systems add more agents, tenants and continuously accessed data sources.

BlueField-4 combines a 64-core Grace CPU, inline acceleration engines, PCIe Gen6 host connectivity and an 800 Gb/s network interface. NVIDIA says this design allows software to set policy and provisioning while hardware handles networking, storage and security locally, without sending the work back to the host CPU.

DOCA is the software bridge in the design. NVIDIA says its containerized microservices can run directly on BlueField-4, while its libraries and SDKs expose accelerated capabilities for networking, security, storage, telemetry, provisioning and updates.

The operational targets

• Virtual private clouds: NVIDIA identifies isolated tenant and application traffic on shared AI infrastructure as a Scale-In use case.
• Security: NVIDIA says BlueField-4 can enforce network, file and object access controls outside the host operating system.
• Storage: BlueField-4 is designed to accelerate storage protocols, virtualization and data movement between AI compute and external data systems.

NVIDIA says Spectrum-X Ethernet manages congestion and traffic isolation across the Scale-In path. The company also says its BlueField-4 and Spectrum-X storage path can deliver up to 1.45 times the throughput of off-the-shelf Ethernet. These are company-supplied architecture and performance claims; the post provides no independent validation or production deployment results.

NVIDIA says Scale-In is co-designed with Vera Rubin infrastructure so processor performance, memory, PCIe, networking, acceleration and software can scale together. The proposition is broader than a faster network card: AI factories will need the systems feeding and governing GPUs to keep pace with the GPUs themselves.