AIPS BOOM

NVIDIA-BlueField-DPU

7 min read

NVIDIA BlueField DPU

Type: Technology Tags: NVIDIA, DPU, data processing unit, networking, offload, security, SmartNIC, BlueField, infrastructure Related: NVIDIA-DOCA, NVIDIA-DOCA-OFED, DOCA-DPA, DOCA-Flow, DOCA-RDMA, DOCA-GPUNetIO, DOCA-PCC, DOCA-Telemetry-Service, DOCA-App-Shield, DOCA-Device-Emulation, DOCA-SNAP, OVS-DOCA, NVIDIA-Firmware-Tools, NVIDIA-BlueField-4, NVIDIA-DOCA-Platform-Framework, NVIDIA-NCX-Infra-Controller, NVIDIA-STX, NVIDIA-CMX, NVIDIA-Vera-Rubin-POD, NVIDIA-Spectrum-6-SPX, NVIDIA-AI-Data-Platform, NVIDIA-Enterprise-Reference-Architectures, NVIDIA-RTX-PRO-AI-Factory, NVIDIA-HGX-AI-Factory, NVIDIA-NVL72-AI-Factory, NVIDIA-DGX-BasePOD-B200-H200-H100-RA, NVIDIA-DGX-SuperPOD-B200-RA, NVIDIA-DGX-SuperPOD-GB200-RA, NVIDIA-ConnectX-InfiniBand, NVIDIA-ConnectX-9, GPUDirect-RDMA, NVIDIA-Network-Operator, NVIDIA-Rivermax, NVIDIA-AI-Enterprise, NVIDIA-DCGM, NVIDIA-GPU-Operator Sources: NVIDIA official documentation (live fetch attempted 2026-04-10; updated from https://docs.nvidia.com/doca/sdk/index.html, https://developer.nvidia.com/networking/doca, https://docs.nvidia.com/networking/display/bluefieldbsp4140/troubleshooting-and-how-tos, https://docs.nvidia.com/networking/display/kubernetes2610/index.html, https://www.nvidia.com/en-us/data-center/ai-storage/stx/, https://www.nvidia.com/en-us/data-center/ai-storage/cmx/, https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/, https://www.nvidia.com/en-us/networking/products/data-processing-unit/) Last Updated: 2026-05-09

Summary

NVIDIA BlueField is a Data Processing Unit (DPU) — a programmable network adapter that combines a high-speed ConnectX NIC with an embedded ARM multicore CPU, memory, and hardware accelerators for networking, security, and storage tasks. BlueField offloads infrastructure services (firewall, load balancer, storage compression/encryption, RDMA routing, telemetry) from server CPUs to the DPU, freeing server CPUs entirely for application workloads while providing stronger security isolation. In AI data centers, BlueField-3 offloads NCCL networking, storage virtualization, and tenant isolation, enabling GPU servers to run more AI compute per server by reclaiming infrastructure CPU cycles.

Detail

Purpose

Modern data center servers spend 20–30% of server CPU cycles on infrastructure functions: network packet processing (firewall, routing, QoS), storage virtualization (encryption, compression, deduplication), and security monitoring. This “infrastructure tax” reduces the CPUs available for the revenue-generating application (AI training, database queries, web serving). BlueField DPU moves these functions off the server CPU onto the DPU’s embedded ARM cores — which are purpose-built for network packet processing — returning those CPU cycles to applications. For GPU AI servers, the benefit is even greater: every freed CPU cycle can be used for data preprocessing or orchestration that accelerates GPU utilization.

Key Features

BlueField Generations:

DPUConnectX NICEmbedded CPUSpeedKey Features
BlueField-1ConnectX-516× ARM A72100 Gb/sInitial DPU; storage/networking offload
BlueField-2ConnectX-6 Dx8× ARM A72200 Gb/sFull data center offload; OVS-DPDK; RegEx engine
BlueField-3ConnectX-716× ARM A78400 Gb/sAI-optimized DPU; DOCA Telemetry; NCCL offload; PCIe Gen5
BlueField-4Next-gen NVIDIA networkingPublic STX/CMX materialAI data/storage classNVIDIA-STX, NVIDIA-CMX, context memory, AI-native storage

Core Capabilities:

  • Networking Offloads: Open vSwitch through OVS-DOCA, VXLAN/Geneve tunneling, BGP routing, firewall (conntrack, ACL), rate limiting, and hardware packet steering through DOCA-Flow
  • Storage Offloads: NVMe-oF target (GPU-Direct Storage integration), RAID, compression (LZ4, zstd hardware engine on BF-3), encryption (AES-256 hardware engine)
  • Security: Hardware-enforced tenant isolation; cryptographic attestation of DPU and host; secure boot chain independent of server OS; DPU-side host introspection through DOCA-App-Shield
  • Telemetry Engine (BF-3): BlueField-3 includes a telemetry accelerator for packet analysis and metrics collection; DOCA-Telemetry-Service exposes BlueField/host/network metrics to observability stacks
  • SNAP (Software-Defined NVMe/VirtIO): DOCA-SNAP emulates NVMe, virtio-blk, and virtio-fs devices to hosts and VMs; enables storage virtualization with a DPU-managed backend
  • Device Emulation: DOCA-Device-Emulation provides lower-level APIs for emulating host-facing PCIe devices from BlueField software
  • DOCA SDK: NVIDIA-DOCA is the software framework for programming BlueField and SuperNIC infrastructure services; current DOCA docs include SDK, runtime, DOCA-Host, DOCA-OFED, reference applications, tools, services, driver documentation, and libraries/services such as DOCA-DPA, DOCA-Flow, DOCA-RDMA, DOCA-GPUNetIO, DOCA-PCC, DOCA-Telemetry-Service, DOCA-App-Shield, DOCA-Device-Emulation, DOCA-SNAP, and OVS-DOCA
  • Arm Neoverse N2 Cores (BF-3): 16× ARM Neoverse N2 @ 2 GHz; 16+ GB LPDDR5 on-DPU memory; runs a full Ubuntu or Wind River Linux OS independently of server host
  • NCCL Offload (BF-3): BlueField-3 + DOCA-NCCL enables in-network AllReduce and collective offloading for AI training; reduces server CPU burden for communication coordination
  • AI-native storage direction: current NVIDIA STX and CMX pages position BlueField-4-class DPUs as storage/context processors for NVIDIA-STX, NVIDIA-CMX, and NVIDIA-AI-Data-Platform designs.
  • Vera Rubin POD direction: current Vera Rubin POD material places BlueField-4 STX racks and CMX context memory storage beside Rubin compute, LPX inference, and SPX networking.

Use Cases

  • AI Training Infrastructure: Offload networking and storage from CPU in DGX nodes; BlueField-3 handles NCCL network communication processing, freeing CPUs for data preprocessing
  • Secure Multi-Tenant Cloud: Hyperscale clouds use BlueField to implement network isolation between tenants in hardware — firewall, QoS, VPC routing all run on DPU, not server CPU
  • Storage Virtualization (Disaggregated Storage): BlueField presents NVMe-oF targets to servers; enables software-defined storage with hardware-grade performance
  • Zero-Trust Security: BlueField’s independent OS enables a “guardian” that monitors server behavior from outside the main OS; detects and responds to host compromise
  • 5G and Telecom: BlueField-based SmartNIC for vRAN (virtual Radio Access Network) packet processing and timing
  • NVIDIA SuperPOD: BlueField-3 in DGX H100 nodes handles storage (NVMe-oF to VAST/WekaFS) and provides tenant isolation

Hardware Requirements / Compatibility

  • Host Interface: PCIe Gen4 x16 (BF-2); PCIe Gen5 x16 (BF-3)
  • OS (DPU): Ubuntu 22.04 aarch64 on embedded ARM; Wind River Linux (real-time option); DOCA runs on DPU OS
  • OS (Host): Linux (Ubuntu/RHEL) for full DOCA integration; Windows Server (limited)
  • Power: 25–75W TDP (model-dependent)
  • Driver: NVIDIA-DOCA-OFED / DOCA-Host on current Linux hosts; legacy NVIDIA-MLNX-OFED for older supported deployments; DOCA runtime on the DPU OS

Language Bindings / APIs

  • NVIDIA DOCA SDK: NVIDIA-DOCA provides APIs, libraries, tools, services, drivers, and examples for packet processing, RDMA, DPA, storage, crypto, telemetry, management, and reference applications
  • Host driver stack: NVIDIA-DOCA-OFED is the current Linux host-driver profile for NVIDIA networking; BlueField-3 platforms must use DOCA-Host.
  • DPDK: BlueField supports DPDK for high-performance user-space networking
  • Kubernetes: NVIDIA-Network-Operator manages NVIDIA networking components in Kubernetes clusters, including DOCA-OFED driver containers and RDMA/GPUDirect networking resources
  • CLI: NVIDIA-Firmware-Tools commands such as mlxconfig and mst for device management; bfb-install for DPU firmware and OS flashing

Connections

  • NVIDIA-DOCA — DOCA is the primary SDK, runtime, driver, and services framework for BlueField infrastructure offload.
  • NVIDIA-DOCA-OFED — current DOCA-Host Linux driver profile for BlueField/ConnectX host systems.
  • DOCA-DPA — DPA programs communication-centric offloads on BlueField hardware.
  • DOCA-Flow — Flow builds hardware-accelerated packet-processing pipes on BlueField.
  • DOCA-RDMA — RDMA API for memory access over InfiniBand/RoCE on BlueField/ConnectX.
  • DOCA-GPUNetIO — GPU-centric packet processing that can use BlueField DMA and GPUDirect paths.
  • DOCA-PCC — programmable congestion-control API accelerated by BlueField-3.
  • DOCA-Telemetry-Service — DOCA service for collecting and exporting BlueField, host, network, RDMA, and DCGM-adjacent metrics.
  • DOCA-App-Shield — DPU-side host and VM introspection API for runtime security and forensics.
  • DOCA-Device-Emulation — PCIe device emulation framework for host-facing BlueField services.
  • DOCA-SNAP — BlueField storage virtualization services for NVMe, virtio-blk, and virtio-fs.
  • OVS-DOCA — DOCA-backed Open vSwitch datapath offload for virtualized networking.
  • NVIDIA-Firmware-Tools — MFT provides low-level firmware and device configuration tools such as mst and mlxconfig.
  • NVIDIA-BlueField-4 — generation-specific page for current STX/CMX and AI data platform DPU direction.
  • NVIDIA-DOCA-Platform-Framework — DPF provisions and orchestrates BlueField DPUs and DPU services in cloud environments.
  • NVIDIA-NCX-Infra-Controller — NCX Infra Controller uses DPU-enforced isolation for secure bare-metal lifecycle management.
  • NVIDIA-STX — STX uses BlueField-4-class acceleration for AI-native storage and data platform designs.
  • NVIDIA-CMX — CMX uses BlueField-4 for context memory storage, NVMe services, KV-cache integrity, and encryption.
  • NVIDIA-Vera-Rubin-POD — BlueField-4 STX racks host context memory storage in the POD-scale architecture.
  • NVIDIA-Spectrum-6-SPX — networking rack carrying traffic between compute, inference, and BlueField-4 STX storage racks.
  • NVIDIA-AI-Data-Platform — AI Data Platform uses BlueField to accelerate data access across enterprise storage paths.
  • NVIDIA-Enterprise-Reference-Architectures — current Enterprise RAs use BlueField DPUs and SuperNIC roles for AI factory fabrics.
  • NVIDIA-RTX-PRO-AI-Factory, NVIDIA-HGX-AI-Factory, and NVIDIA-NVL72-AI-Factory — hardware RAs that include BlueField in compute, converged, or infrastructure roles.
  • NVIDIA-DGX-BasePOD-B200-H200-H100-RA, NVIDIA-DGX-SuperPOD-B200-RA, and NVIDIA-DGX-SuperPOD-GB200-RA — DGX RAs where BlueField supports storage, management, or front-end networking roles.
  • NVIDIA-ConnectX-InfiniBand — BlueField is a ConnectX HCA with an embedded ARM processor; shares ASIC design; BF-3 uses ConnectX-7 NIC silicon
  • NVIDIA-ConnectX-9 — adjacent next-generation SuperNIC direction for 1.6 Tb/s-class AI networking.
  • GPUDirect-RDMA — BlueField supports GPUDirect RDMA; BF-3 NCCL offload enables GPU communication coordinated by DPU
  • NVIDIA-Network-Operator — Kubernetes operator that deploys DOCA-OFED and RDMA networking resources for BlueField/ConnectX environments.
  • NVIDIA-Rivermax — Rivermax supports BlueField DPUs for timing-sensitive streaming and DOCA Rivermax workflows.
  • NVIDIA-AI-Enterprise — BlueField DPUs are part of NVIDIA’s AI infrastructure stack; supported under AI Enterprise for security and networking offload in AI data centers
  • NVIDIA-DCGM — BlueField’s telemetry engine and DOCA can feed network metrics to DCGM-based monitoring for holistic AI cluster observability
  • NVIDIA-GPU-Operator — GPU Operator works with Network Operator in Kubernetes clusters that need both GPUs and accelerated networking

Resources

Graph View

Backlinks