BlueField-3 DPU in GPU servers: what a DPU is, what it offloads and when you need one

What a BlueField-3 DPU is

A BlueField-3 DPU is a network adapter that carries its own computer. Next to two ports of up to 400 Gb/s, the card holds 16 Arm cores, 32 GB of memory, an SSD and its own operating system, and it runs network, storage and security functions there instead of on the server's processors. In our judgement a single inference server does not need one. It becomes useful where several tenants share GPU servers, or where storage and network services have to run outside the host's operating system.

NVIDIA's BlueField page of 21 October 2026 answers the question of what a DPU is this way: "A DPU is a programmable infrastructure processor that combines a high-performance CPU" with a high-speed network interface and hardware acceleration engines, built to offload, accelerate and isolate infrastructure services from the host. The same page calls the BlueField-3 DPU "a 400 Gb/s infrastructure compute platform".

BlueField-3 specs: Arm cores, memory, ports

NVIDIA BlueField-3 Networking Platform User Guide, index and Introduction (10 November 2026) and Specifications (23 July 2026); NVIDIA BlueField-3 datasheet 5194714, Chinese edition (May 2026); read 12 November 2026.

BlueField-3 DPU, BlueField-3 SuperNIC and ConnectX-7

"BlueField-3" is the name of two hardware classes, and the 16 Arm cores apply to one of them only.

NVIDIA BlueField-3 Networking Platform User Guide, index (10 November 2026) and Hardware Installation (23 July 2026); DOCA 3.5.0, BlueField Modes of Operation (18 September 2026); ConnectX-7 Adapter Cards User Manual (12 August 2026) and datasheet; NVIDIA's developer blog of 18 May 2025; NVIDIA's HGX and RTX PRO AI Factory reference architectures (18 May 2026); read 12 November 2026.

A ConnectX-7 is the adapter without the computer, and in NIC mode a BlueField-3 behaves like one: NVIDIA's DOCA documentation says that "the device functions as an NVIDIA® ConnectX® network adapter", and on the BlueField-3 the Arm cores are then inactive. NVIDIA's product page for RTX PRO Servers of 11 June 2026 lists its 2U design for two RTX PRO 6000 Blackwell Server Edition cards with ConnectX-7 or a BlueField-3 SuperNIC, so the two are alternatives for traffic between servers. The faster adapter of the Blackwell generation is described in our article on the ConnectX-8 SuperNIC.

NVIDIA BlueField-3 Networking Platform User Guide, index (10 November 2026), models in mass production with integrated BMC; read 12 November 2026.

Three further models stand in the guide's end-of-life table and are named here only so that they are recognised in an old parts list: the B3210 (DPU, 100GbE), the B3140L (SuperNIC, one 400GbE port) and the B3240 for liquid-cooled servers. The single-slot B3240 ending in PA0 is, by the guide, for cold-aisle servers only. For the B3240 the guide gives InfiniBand NDR as the default link type, while NVIDIA's HGX reference architecture writes "400GbE (default mode) /NDR IB"; state the link type you need in the order.

In the servers we build to order, the network cards are part of the configuration. Name the model, the link type and the server you plan, and we reply with a written quote for the configuration within one business day.

DPU mode, Zero Trust mode and NIC mode

What the card does depends on its operating mode, which NVIDIA's DOCA 3.5.0 documentation of 18 September 2026 describes.

• DPU mode. In NVIDIA's words, "the NIC resources and functionality are owned and controlled by the embedded Arm subsystem", and "All network communication to the host flows through a virtual switch control plane hosted on the Arm cores."
• Zero Trust mode. A variant of DPU mode in which the administrator of the host cannot manage the BlueField from the host; port ownership, hardware counters, tracing, RShim and firmware flashing can be restricted.
• NIC mode. The card works as a ConnectX adapter. NVIDIA names reduced power consumption, improved network performance and a smaller host memory footprint as the benefits.

The factory setting follows the hardware class: "BlueField-2 DPU and BlueField-3 DPU SKUs are shipped in DPU Mode by default", and "BlueField-3 SuperNIC SKUs are shipped in NIC Mode by default." The mode is changed with the mlxconfig parameter INTERNAL_CPU_OFFLOAD_ENGINE and a reset or power cycle; a direct change from Zero Trust mode to NIC mode is not supported. Two more limits apply: Socket Direct is not supported in DPU mode, and Multi-Host is not supported in NIC mode.

What a DPU takes off the host: switching, storage, security

The software is NVIDIA DOCA, current version 3.5.0. It has two parts: DOCA-Host, installed on the server, and the BlueField bundle on the Arm cores, whose default operating system in this release is Ubuntu 24.04. The host needs DOCA-Host, because NVIDIA's release notes of 23 October 2026 state that BlueField-3 devices are not supported with MLNX_OFED as the host driver.

NVIDIA DOCA 3.5.0, DOCA Services (1 September 2026), read 12 November 2026.

For encryption the user guide lists inline IPsec and TLS, MACsec, AES-XTS block encryption and public key acceleration. All of these services run on the card, outside the host's operating system, and the host sees a network device or an NVMe drive.

Where NVIDIA's reference designs place BlueField-3

NVIDIA DGX B300 User Guide, introduction; NVIDIA HGX AI Factory and RTX PRO AI Factory enterprise reference architectures, Components and appendices (18 May 2026); NVIDIA RTX PRO Servers product page (11 June 2026); read 12 November 2026.

North-south is the traffic in and out of the cluster: users, storage and management. East-west is the traffic between GPU servers. The HGX architecture states that the "NVIDIA BlueField-3 DPU is optimized for the North/South network" and names its tasks there: a control-plane platform for provisioning, storage acceleration with BlueField SNAP, a zero-trust mode that works independently of the host with firewall and micro-segmentation services, and the card's own BMC with Redfish on a dedicated 1GbE port. For HGX B300 platforms NVIDIA recommends the B3240 "rather than the NVIDIA BlueField-3 B3220 that is common in NVIDIA HGX H100/200 and HGX B200 systems".

NVIDIA's own DGX system sets its DPUs to the simpler mode. The user guide says: "The default mode for BlueField-3 DPUs in DGX systems is NIC mode." As delivered, the two DPUs described in our DGX B300 specifications therefore work as network adapters. Server makers adopt the reference designs: Gigabyte lists one BlueField-3 DPU for its XL44-SX2-AAS1, covered in Gigabyte servers for the RTX PRO 6000 Server Edition, and NVIDIA's B3240 recommendation is the reference for HGX B300 servers such as the Supermicro SYS-822GS-NB3RT.

When a GPU server needs a DPU: three cases

our assessment, based on NVIDIA's configuration guide for NVIDIA-Certified Systems (read 12 November 2026), DOCA 3.5.0 documentation and the reference architectures of 18 May 2026.

The three cases below are our assessment, drawn from the documents named under the table.

A single inference server: no. NVIDIA's configuration guide for NVIDIA-Certified Systems says: "Single node deployments typically do not require high-speed networking to connect multiple nodes for your AI workload." One team administers the host, so there is nobody to isolate the network services from, and a card in DPU mode would be a second computer to patch, with its own operating system, firmware and BMC. This is our judgement; NVIDIA's reference architectures, which describe clusters, include one in every RTX PRO Server and HGX B300 server.

Two to four servers of one operator: usually a standard NIC. Servers that run complete model copies behind a load balancer need ordinary Ethernet, and our article on a GPU inference cluster without InfiniBand plans two 25 GbE ports per server. When one model spans servers, the east-west network needs RDMA at 200 to 400 Gb/s. That is the job of ConnectX adapters or of BlueField-3 SuperNICs in NIC mode.

Several tenants or network storage: yes. When departments or customers get whole servers with administrator rights, rules enforced inside the host can be changed by the tenant. On a DPU in Zero Trust mode, the virtual switch, routing and the L4 firewall run on the card, with the switch's control plane on the Arm cores, out of the tenant's reach; our article on network segmentation and micro­segmentation explains the principle. The storage case is similar: with SNAP, volumes of a networked storage system appear to the host as local NVMe or virtio-blk devices. In clusters run on Kubernetes, NVIDIA's DOCA Platform Framework provisions the DPUs.

Sizing is free of charge with us. Tell us how many servers you plan, who administers them and where the storage sits, and our team proposes a configuration with the adapters for each network.

What the server must provide: power cable, slot, airflow

A DPU asks more of the chassis than a plain adapter. The user guide states that the "B3240, B3220, B3210 and B3210E DPUs require a supplementary 8-pin ATX power supply connectivity", and explains: "While the PCIe slot feeds 66W, the 8-pin PCIe ATX power supply cable provides additional power." Without the cable "the DPU will not complete the power on procedure", and the cable is not in the package. NVIDIA adds two warnings: the connector "is not compatible with an EPS12V power cable source", and "Do not link the CPU power cable to the BlueField-3 DPU PCIe ATX power connector, as their pin configurations differ."

The SuperNICs take their power from the slot, which must supply at least 75 W. A wattage for the cards themselves is not public: the guide points to product specifications that NVIDIA provides through NVOnline or a representative, and the HGX architecture only notes that "Some BlueField-3 configurations for north-south connectivity can draw more than 75 W".

Three more points belong in the chassis check. The standard B3240 takes two slots at full height and half length (111.15 × 167.65 mm), while the B3140H is half height (68.90 × 167.65 mm). The cards need an airflow of 200 LFM at the maximum ambient temperature of 55 °C, in one direction, "from the heatsink to the network ports"; reversed airflow has to be validated thermally. And the reference architecture asks for "One Gen5 x16 link per DPU, SuperNIC or adapter." On the network side, NVIDIA's DGX SuperPOD design joins the BMCs of the BlueField cards to the out-of-band network, where our guide to securing a GPU server also puts the server's BMC.

vSphere and Kubernetes: what is documented

On VMware, the function is called vSphere Distributed Services Engine. Broadcom's documentation for vSphere 8.0 (page of 28 October 2026) and 9.0 describes it as "a core vSphere capability that enables customers to use DPUs with vSphere", "available on servers with pre-installed DPUs" from NVIDIA and AMD in server designs from Dell, HPE, Lenovo and Fujitsu. It offloads networking: "Traffic forwarding logic is offloaded from the ESXi host to the vSphere Distributed Switch backed by the DPU." Storage offload and bare-metal management are "currently not supported". Broadcom states that the function "does not require a separate ESXi license". Its release notes name BlueField-2 DPUs for server designs from Lenovo, Dell and Fujitsu, and BlueField-3 is not named on the Broadcom pages we read, so support for a given server has to be checked in the Broadcom Compatibility Guide. On vSphere the DPU therefore comes pre-installed with the server model. The host side of GPU servers on vSphere is in our guide to GPU servers in a vSphere cluster and in NVIDIA AI Enterprise on vSphere.

For Kubernetes, NVIDIA's DOCA Platform Framework 26.8.0 "provisions and orchestrates NVIDIA BlueField DPUs in cloud environments". Its platform page of 14 October 2026 lists upstream Kubernetes 1.33 to 1.36, Ubuntu 24.04 LTS as the only tested host system and dual-port BlueField-3 DPUs, with the B3240, the B3220 and the B3210 as recommended models (the user guide lists the B3210 as end of life); Socket Direct is not supported. Red Hat OpenShift is not on that page. The cluster layer itself is covered in private LLMs on Kubernetes and, for scheduling between teams, in NVIDIA Run:ai licensing and hardware requirements.

BlueField-4: what NVIDIA has announced

All statements on BlueField-4 are announcements. NVIDIA's BlueField page of 21 October 2026 writes: "The NVIDIA BlueField-4 DPU is an 800 Gb/s infrastructure platform for gigascale AI factories." NVIDIA's blog on the Vera Rubin platform (5 January 2026, modified 21 August 2026) describes a package of a 64-core NVIDIA Grace CPU and an integrated ConnectX-9 and sets it against BlueField-3: 800 instead of 400 Gb/s, 64 Neoverse V2 cores instead of 16 A78 cores, 128 GB of memory instead of 32 GB. The press release of 5 January 2026 counts BlueField-4 among the six chips of the Rubin platform and names the second half of 2026 for Rubin-based products from partners. On 12 November 2026, NVIDIA's documentation hub listed hardware guides for BlueField-3 and BlueField-2 only, and DOCA 3.5.0 listed no BlueField-4 firmware. The reference architectures cited above specify BlueField-3.

We supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project. Send us the reference design or the parts list you work from, and we check which BlueField-3 model, power cable and slot the chosen chassis provides.

FAQ