Networking NVIDIA BlueField-4 Networking Platform User Guide

Introduction

The NVIDIA® BlueField®-4 data processing unit (DPU) is an accelerated infrastructure platform purpose-built for gigascale AI factories, delivering throughput up to 800 gigabits per second (Gb/s). By combining the NVIDIA Grace™ CPU for high-performance computing with NVIDIA® ConnectX®-9 networking, BlueField-4 unlocks new levels of acceleration, efficiency, and cybersecurity, powering every layer of AI data center infrastructure

he platform integrates powerful data processing, native service function chaining, built-in security acceleration, and massive I/O performance. BlueField drives efficiency across the data center as a data processing unit in AI compute nodes and as a high-performance storage controller that enables accelerated, secure, and software-defined access to data. AI-native organizations, service providers, and enterprises can now build multi-tenant AI infrastructure and data platforms with greater scale, achieving peak workload and power efficiency while supporting high-performance storage, rapid data access, and large-scale inference.

Accelerating Infrastructure for AI 

BlueField-4 is an open, programmable platform powered by NVIDIA DOCA™, designed to accelerate, protect, and scale modern AI infrastructure. It natively runs DOCA microservices, delivering prebuilt, containerized applications for AI networking, orchestration, real-time threat detection, and data storage acceleration with higher efficiency and throughput.
The BlueField architecture spans the full NVIDIA AI infrastructure portfolio—from computing systems to scale-out data center designs—ensuring consistent performance, efficiency, and security at every level. Enterprises can further extend functionality by leveraging a broad range of accelerated infrastructure and cybersecurity applications from leading software providers.

BlueField Advanced Secure Trusted Resource Architecture (ASTRA)

The BlueField-4 platform introduces the Advanced Secure Trusted Resource Architecture (ASTRA), designed to extend security and manageability in the NVIDIA Rubin platform with NVIDIA ConnectX SuperNIC™ cards, enabling secure bare-metal compute instances with zero-trust tenant isolation and full software-defined infrastructure control.

ASTRA extends provisioning, manageability, and policy enforcement across the East-West fabric. The BlueField DPU controls all network I/O to and from the compute node via dedicated links between the BlueField-4 DPU and NVIDIA ConnectX-9 SuperNICs. By isolating the SuperNIC control plane from the host operating system, BlueField ASTRA prevents tenant workloads, even on bare metal, from accessing or interfering with network configuration and provisioning.

Cooling Mechanisms

The BlueField-4 DPUs provides server flexibility with support for air and liquid cooled solutions.

The BlueField-4 Liquid Cooled DPU is a high-performance data processing unit designed to seamlessly integrate into fully liquid-cooled server platforms. Offering the same advanced capabilities and features as the standard BlueField-4, it ensures optimal performance and efficiency in dense, thermally demanding data center environments. This liquid-cooled variant enables 100% liquid-cooled server deployments, supporting modern, sustainable infrastructure needs without compromising functionality or compatibility.

System Requirements

Item

Description

PCI Express slot

PCIe Gen6 (64GT/s) through x16 edge connector (PCIe bifurcation two x8 PCIe lanes - Socket Direct enabled by the Golden Finger pinouts)

System Power Supply

A minimum of 300W via the 12-pin power supply connectivity through the external power supply connector (CEM6).

The power supply harness is included in the package. 

Refer to the Hardware Installation for important notes and warnings on powering up and down the card. 

Operating System

BlueField-4 platforms is shipped with Ubuntu – a Linux commercial operating system – which includes the DOCA-HOST, and is capable of running all customer-based Linux applications seamlessly. For more information, please refer to the DOCA SDK documentation or NVIDIA BlueField BSP Documentation.

Connectivity

  • Interoperable with 100/200/400/800 Gb/s Ethernet switches and EDR/HDR100/HDR/NDR200/NDR/XDR InfiniBand switches

  • Powered connectors for optical and active cable support

For detailed information, see Specifications.

Package Contents 

Prior to unpacking your product, it is important to make sure your server meets all the system requirements listed above for a smooth installation. Be sure to inspect each piece of equipment shipped in the packing box. If anything is missing or damaged, contact your reseller.

Card Package

Item

Description

Card

1x BlueField-4 DPU

Power Supply Harness

[Only in B4240] 1 x Black power harness that connects four female PCIe 8-pin (2x4) power connectors to a single male PCIe CEM5 16-pin power connector


Accessories Kit 

This is an optional accessories kit used for debugging purposes and can be ordered separately. 

Optional Accessories, not included in the package, can be purchased separately:

OPN

Description

930-9XCBL-000A-000

NVIDIA 200mm Cable Extender for Low-Speed Signals Over 30p Debug Connector

Features and Benefits

This section describes hardware features and capabilities. Please refer to the relevant driver and/or firmware release notes for feature availability.

Feature

Description

InfiniBand Architecture Specification v1.5 and 1.7 compliant

BlueField-4 platforms deliver low latency, high bandwidth, and computing efficiency for high-performance computing (HPC), artificial intelligence (AI), and hyperscale cloud data centers applications.
BlueField-4 platforms are InfiniBand Architecture Specification v1.5 and 1.7 compliant.

InfiniBand Network Protocols and Rates: 

Protocol


Standard

Port Width Rate (Gb/s)

Encoding

4x (4 lane)
Port

2x (2 lane)
Port

1x (1 lane)
Port

NDR

IBTA Vol2 1.5

425

212.5

--

PAM4

HDR

IBTA Vol2 1.4

212.5

106.25

53.125

PAM4

  With these basic rates, 800Gb/s can be built from 4 ports of 200Gb/s, for example.

Up to 400 Gigabit Ethernet




BlueField-4 platforms comply with the following IEEE 802.3 standards:
800GbE / 400GbE / 200GbE / 100GbE / 50GbE 

Protocol

Ethernet Network Rate

IEEE 802.3dj

  • Clause 179

  • Annex 176D

800/400/200 Gigabit Ethernet

IEEE 802.3df

  • Clause 162

  • Annex 120G

800/400/200/100 Gigabit Ethernet

IEEE 802.3-2022

  • Clause 136

  • Annex 120E

  • Annex 135F

Ethernet Technology Consortium

400/200/100/50Gb/s Gigabit Ethernet

IEEE 802.3-2022

  • Clause 73

Auto-negotiation

IEEE 802.3-2022

  • Section 136.8.11

IEEE 802.3df

  • Section 162.8.11

IEEE 802.3dj

  • Annex 178B

Link training

IEEE 802.1AX 

Link Aggregation

IEEE 802.1Q
IEEE 802.1P VLAN tags and priority


IEEE 802.1Qau (QCN)
Congestion Notification
IEEE 802.1Qaz (ETS)
EEE 802.1Qbb (PFC)
IEEE 802.1Qbg
IEEE 1588v2
Jumbo frame support (9.6KB)


On-board Memory

  • Quad SPI NOR FLASH - includes 256Mbit for Firmware image.

  • UVPS EEPROM - includes 2Mbit.

  • FRU EEPROM - Stores the parameters and personality of the card. The EEPROM capacity is 128Kbit. FRU I2C address is (0x50) and is accessible through the PCIe SMBus.

  • DPU_BMC Flashes: 

    • 2x 64MByte for BMC Image

    • 512MByte for Config Data

  • SSD (onboard BGA) 512GB in M.2 detachable device for user SoC OS, logs and application SW.

  • 8x LPDDR5x (organized as 16 channels)- total capacity 128GB @8533MT/s with 256-bit bus width

BlueField-4 IC

The BlueField-4 platforms integrate a 64 Arm Core Neoverse v2, SPECInt2K17 – 250

Overlay Networks

In order to better scale their networks, data center operators often create overlay networks that carry traffic from individual virtual machines over logical tunnels in encapsulated formats such as NVGRE and VXLAN. While this solves network scalability issues, it hides the TCP packet from the hardware offloading engines, placing higher loads on the host CPU. BlueField-4 platforms effectively addresses this by providing advanced NVGRE and VXLAN hardware offloading engines that encapsulate and de-capsulate the overlay protocol.

RDMA and RDMA over Converged InfiniBand/Ethernet (RoCE)

Utilizing IBTA RDMA (Remote Data Memory Access) and RoCE (RDMA over Converged InfiniBand/Ethernet) technology, the BlueField-4 platforms deliver low-latency and high-performance over InfiniBand/Ethernet networks. Leveraging data center bridging (DCB) capabilities as well as advanced congestion control hardware mechanisms, RoCE provides efficient low-latency RDMA services over Layer 2 and Layer 3 networks.

Quality of Service (QoS)

Support for port-based Quality of Service enabling various application requirements for latency and SLA.

Storage Acceleration

  • A consolidated compute and storage network achieves significant cost-performance advantages over multi-fabric networks. Standard block and file access protocols can leverage RDMA for high-performance storage access: NVMe over Fabric offloads for the target machine

  • BlueField-4 cards may operate as a co-processor offloading specific storage tasks from the host, isolating part of the storage media from the host, or enabling abstraction of software-defined storage logic using the NVIDIA BlueField-4 Arm cores. On the storage initiator side, BlueField-4 networking platforms can prove an efficient solution for hyper-converged systems to enable the host CPU to focus on compute while all the storage interface is handled through the Arm cores.

NVMe-oF

Non-volatile Memory Express (NVMe) over Fabrics is a protocol for communicating block storage IO requests over RDMA to transfer data between a host computer and a target solid-state storage device or system over a network. BlueField-4 platforms may operate as a co-processor offloading specific storage tasks from the host using its powerful NVMe over Fabrics Offload accelerator.

SR-IOV

The SR-IOV technology provides dedicated adapter resources and guaranteed isolation and protection for virtual machines (VM) within the server.

High-Performance Accelerations

  • Tag Matching and Rendezvous Offloads

  • Adaptive Routing on Reliable Transport

  • Burst Buffer Offloads for Background Checkpointing

GPU Direct

GPUDirect RDMA is a technology that provides a direct P2P (Peer-to-Peer) data path between the GPU Memory directly to/from the NVIDIA HCA devices. This provides a significant decrease in GPU-GPU communication latency and completely offloads the CPU, removing it from all GPU-GPU communications across the network. BlueField-4 platforms use high-speed DMA transfers to copy data between P2P devices resulting in more efficient system applications

Isolation

BlueField-4 platforms function as a “computer-in-front-of-a-computer,” unlocking unlimited opportunities for custom security applications on its Arm processors, fully isolated from the host’s CPU. In the event of a compromised host, BlueField-4 may detect/block malicious activities in real-time and at wire speed to prevent the attack from spreading further.

Cryptography Accelerations

From IPsec and TLS data-in-motion inline encryption to AES-XTS block-level data-at-rest encryption and public key acceleration, BlueField-4 hardware-based accelerations offload the crypto operations and free up the CPU, reducing latency and enabling scalable crypto solutions. BlueField-4 “host-unaware” solutions may transmit and receive data, while BlueField-4 acts as a bump-in-the-wire for crypto.

Security Accelerators

A consolidated compute and network solution based on BlueField-4 achieves significant advantages over a centralized security server solution. Standard encryption protocols and security applications can leverage BlueField-4 compute capabilities and network offloads for security application solutions such as Layer4 Statefull Firewall.

Virtualized Cloud

By leveraging BlueField-4 virtualization offloads, data center administrators can benefit from better server utilization, allowing more virtual machines and more tenants on the same hardware, while reducing the TCO and power consumption

Out-of-Band Management

The BlueField-4 platforms incorporate a 1GbE RJ45 out-of-band port that allows the network operator to establish trust boundaries in accessing the management function to apply it to network resources. It can also be used to ensure management connectivity (including the ability to determine the status of any network component) independent of the status of other in-band network components.

BMC

All BlueField-4 platforms incorporate local NIC BMC (Baseboard Management Controller) hardware on the board. The BMC SoC (system on a chip) can utilize either shared or dedicated NICs for remote access. The BMC node enables remote power cycling, board environment monitoring, BlueField-4 chip temperature monitoring, board power and consumption monitoring, and individual interface resets. The BMC also supports the ability to push a bootstream to BlueField-4. Having a trusted on-board BMC that is fully isolated for the host server ensures highest security for the BlueField-4 platforms.

Baremetal East-West Management (BlueField-4 ASTRA)

The BlueField-4 platform introduces the Advanced Secure Trusted Resource Architecture (ASTRA), designed to extend security and manageability in the NVIDIA Rubin platform with NVIDIA ConnectX SuperNIC™ cards, enabling secure bare-metal compute instances with zero-trust tenant isolation and full software-defined infrastructure control.

Last updated: