ConnectX-7 vs BlueField-3: Do Your Servers Need a SmartNIC or a DPU?

A ConnectX-7 is a network adapter with hardware offloads. A BlueField-3 is a network adapter with a small computer on it. That is the whole distinction in one line: the BlueField-3 DPU embeds a ConnectX-7 network engine and adds its own Arm cores, memory and operating system, so it can run infrastructure software (virtual switching, storage networking, firewalling, encryption) separately from the host. Most servers, including most GPU servers, need the first. A smaller set of deployments, usually multi-tenant, security-sensitive or storage-heavy, justify the second. This guide explains where the line falls and which NVIDIA parts sit on each side of it.

What is the difference between a SmartNIC and a DPU?

A SmartNIC offloads defined functions in silicon. ConnectX-7 handles RDMA over InfiniBand and RoCE, GPUDirect RDMA and GPUDirect Storage, stateless TCP and UDP offloads, overlay encapsulation and, on crypto-enabled SKUs, inline encryption. The host CPU still runs the drivers, the virtual switch control plane and the policy that tells the offloads what to do. The card accelerates; the host decides.

A DPU (data processing unit) adds a general-purpose compute complex to that network engine. The BlueField-3 B3220 pairs an integrated ConnectX-7 with 16 Arm A78 cores at around 2.3 GHz and 32GB of DDR5, programmed through NVIDIA's DOCA SDK and runtime. It boots its own Linux-based operating system and can enforce network, storage and security policy that the host operating system cannot see or override. That last property matters most: the DPU becomes a separate trust domain sitting between the server and the network. Side by side, the three NVIDIA adapters most buyers compare look like this:

• Ports: the ConnectX-7 NDR 400G single-port OSFP gives 1x NDR 400Gb/s InfiniBand. The ConnectX-7 NDR200 dual-port QSFP112 gives 2x 200Gb/s, running InfiniBand NDR200 or 200GbE (VPI). The BlueField-3 B3220 also gives 2x 200Gb/s on QSFP112.
• Host interface: all three use PCIe 5.0 x16.
• Form factor: both ConnectX-7 cards are half-height, half-length (HHHL). The B3220 is a full-height, half-length (FHHL) card, so it is not automatically a drop-in replacement in a low-profile slot.
• On-card compute: none available for your own software on ConnectX-7. The B3220 carries 16 Arm cores and 32GB DDR5.
• Acceleration focus: ConnectX-7 targets collectives, MPI tag matching and GPUDirect. BlueField-3 adds SDN, NVMe-oF, line-rate crypto and RegEx on top of the ConnectX-7 feature set.
• Operations: a ConnectX-7 is a driver and a firmware image. A BlueField-3 is a second operating system per server to image, patch and monitor.

ConnectX-7 vs BlueField-3: which do I need?

Choose ConnectX-7 when the job is moving data fast and nothing more. For multi-node training, the single-port 400G OSFP adapter is the standard one-per-GPU part, with hardware acceleration for collectives and all-to-all traffic and GPUDirect RDMA straight into GPU memory. Where you want two ports, or the option to run InfiniBand or Ethernet from the same card, the dual-port NDR200 card is the flexible choice. For general-purpose servers on 10/25GbE, even ConnectX-7 is more than the job needs: a dual-port SFP28 card such as the ConnectX-4 Lx (PCIe 3.0 x8, RoCE v2) covers virtualization and storage networking at that speed. Choose BlueField-3 when you need infrastructure to run somewhere other than the host. The reasons that hold up in practice:

• Multi-tenancy and bare-metal isolation: a tenant can have root on the server and still cannot change network or security policy, because that policy runs on the DPU.
• Storage offload: the DPU handles NVMe-oF storage networking on the card, keeping that work off the host CPU.
• Line-rate encryption and distributed firewalling, enforced on every server without a software agent that a compromised host could disable.
• Host CPU recovery: when virtual switching, overlays and encryption consume a measurable share of expensive host cores, moving them to the DPU returns those cores to workloads.

Buy ConnectX-7 for bandwidth. Buy BlueField-3 when networking, storage or security policy has to run somewhere the host cannot touch. A DPU you never program is an adapter with idle Arm cores.

Do GPU servers need a DPU?

Not for the GPU fabric. Look at how NVIDIA builds its own systems. The DGX H100 is specified with 8x ConnectX-7 400Gb/s InfiniBand adapters plus 2x BlueField-3 DPUs. Each compute tray in a DGX GB200 NVL72 carries 4x ConnectX-7 single-port 400G OSFP adapters, one per GPU, plus 2x dual-port 400G BlueField-3 DPUs. In NVIDIA's reference designs the per-GPU ConnectX-7 adapters form the east-west compute fabric, while the DPUs typically serve storage, in-band management and north-south traffic into the rest of the data center.

So for a single GPU node or a small cluster, the practical rule is: specify one ConnectX-7 per GPU if you are building a multi-node training fabric (eight per 8-GPU node, as on the HGX H100 rail-optimized InfiniBand training node), and add BlueField-3 only on the storage or tenant-facing network if one of the DPU use cases applies. Inference servers behind a load balancer often need neither at 400G and are well served by a dual-port adapter.

There is one exception worth planning for. Some Ethernet AI fabric designs, NVIDIA Spectrum-X among them, put BlueField-3 on the east-west GPU network as well. The Supermicro SYS-821GE-TNHR configuration Nexus Compute lists pairs 8x H200 SXM with 8x 400G BlueField-3 DPUs for Spectrum-X or Quantum-2 fabrics. If your network team is standardizing on Spectrum-X, the adapter choice follows the fabric design, and it should be settled alongside the switch order. Our InfiniBand vs Ethernet interconnect guide covers that fabric decision.

Can a BlueField-3 replace a ConnectX-7 in the same slot?

Check three things before assuming it can. Physically, the B3220 is FHHL where the ConnectX-7 cards are HHHL, so confirm slot height, riser layout and airflow in the chassis. On the network side, the B3220's two ports run at 200Gb/s each, so it is not a like-for-like swap for a single-port 400G compute adapter. On the software side, the host now reaches the network through a device with its own operating system, which changes driver packaging, provisioning and troubleshooting. The network engine is the same ConnectX-7 generation, so raw port speed is not a reason to pick one over the other.

How much extra work is a DPU to run?

Budget for it honestly. Each BlueField-3 needs an OS image and firmware lifecycle, monitoring, and someone who understands DOCA or the platform software built on it. If the virtualization, storage or security product you already run supports DPU offload, much of that work is packaged for you. If it does not, you are signing up for development work before the DPU does anything a ConnectX-7 would not. Confirm DPU support for the exact software version you will deploy before you order.

What should you settle before ordering?

• Which network each port serves: GPU compute fabric, storage, in-band management or tenant traffic.
• Fabric type: InfiniBand, Ethernet with RoCE, or Spectrum-X, because it decides the adapter and the optics.
• Adapter count per node: one per GPU for multi-node training, fewer for inference and single-node work.
• Slot height, riser layout and airflow for FHHL versus HHHL cards.
• Software: whether anything you run today will actually use the DPU's Arm cores.

Nexus Compute supplies ConnectX-7 and BlueField-3 adapters and complete GPU servers configured with either, sourced through authorized distribution and checked against your platform before dispatch. Browse the NVIDIA ConnectX adapter range, or send us the node count, fabric type and workload and we will return a configuration and pricing within 48 business hours via request a quote.