NVIDIA BlueField-4 DPU Doubles Bandwidth, Adds Grace CPU for Agentic AI

NVIDIA's BlueField-4 DPU brings a 64-core Grace CPU, LPDDR5X memory, and PCIe Gen6 interfaces, with up to 800 Gb/s Ethernet or InfiniBand connectivity. The company claims the new DPU delivers up to six times more compute performance, doubles networking bandwidth, quadruples memory capacity, and offers more than three times the memory bandwidth compared to BlueField-3. These specs target agentic AI workloads, which involve multiple model calls, tool calls, KV-cache lookups, and policy checks per request.

Offloading for Lower Latency

BlueField-4 offloads networking, storage, security, telemetry, and control-plane services from host CPUs. By keeping service logic close to data with embedded compute and LPDDR5X, the DPU reduces host CPU overhead. This offloading improves GPU utilization and reduces inference latency, according to NVIDIA. The DPU accelerates flow steering, RDMA transfers, storage commands, encryption, integrity, and policy enforcement. PCIe Gen6, VirtIO, and DOCA SNAP enable host-visible network and storage device models backed by BlueField acceleration.

KV-Cache and Storage

DOCA Memos manage and share KV-cache across compute and storage nodes, addressing the need for KV-cache persistence across steps in long-context inference. The Vera BlueField-4 STX Storage Processor combines a Vera CPU with a ConnectX-9 SuperNIC, offering up to 1.6 Tb/s Spectrum-X Ethernet connectivity. It provides high-performance NVMe storage access and accelerated data movement.

Software and Security

DOCA is the programmable software foundation for BlueField infrastructure services. DOCA Host-Based Networking supplies server-side Layer 3 routing and BGP functionality. BlueField ASTRA enables Spectrum-X zero-trust, multi-tenant bare-metal deployments. DOCA security services provide zero-trust access, policy enforcement, runtime visibility, and isolation.

System Integration

BlueField operates across the entire Vera Rubin platform, which combines Rubin GPUs, Vera CPUs, NVLink, ConnectX-9, and Spectrum-X. The design aims to lower cost per token and increase tokens per watt. Developers can download DOCA and build sample code to start programming BlueField services.