Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time — we’ll notify you if affected. OEM costs and supply tightening — contact us to secure stock and pricing. For projects or bulk orders, choose “Request for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

Networking & Cybersecurity

NIC, SuperNIC or DPU? What BlueField Does and Whether You Need It

Read the bill of materials for a modern AI server and, somewhere below the GPUs, you will find lines for SuperNICs and DPUs. They are not cheap, there are a lot of them, and most buyers could not say what they are for.

The short answer: they are network cards that have grown a brain. Whether you need the brain depends on how big your cluster is and who shares it.

Three kinds of network card

Compare 🔌 Standard NIC ⚡ SuperNIC 🧠 DPU
What it is A card that moves data in and out of the server A network card tuned for GPU-to-GPU traffic A small computer on a card that runs infrastructure software
Main job Connectivity Keeping AI traffic between servers fast and even under load Networking, storage, security and management, taken off the main processors
Traffic it serves Any East-west
Server to server inside the cluster
North-south
In and out of the cluster, plus storage
Typical count in an 8-GPU server One or two One per GPU in reference designs One, sometimes two
Complexity

Sources: MacMyths: what NVIDIA BlueField does for AI servers, RDP: NIC, SuperNIC and DPU roles explained, FiberMall on BlueField-3. The complexity rating is our own.

The DPU: an assistant that handles the housekeeping

A data processing unit is a processor in its own right. NVIDIA's BlueField-3 has 16 Arm cores and network ports of up to 400 Gb/s. Its job is not AI computation. It runs the infrastructure work around it: software-defined networking, storage access, encryption and firewalling, and management.

The argument for it is simple. Every one of those tasks otherwise takes time from the server's main processors. Move them to a dedicated card and the expensive parts of the server spend more of their time on the workload. There is a security argument too: the infrastructure software runs on separate hardware from the tenant's applications, which matters when several organisations share one cluster.

The SuperNIC: a DPU on a diet

In an AI cluster the heaviest traffic is not between the cluster and the outside world. It is between the GPU servers themselves, as they pass results back and forth during training. The SuperNIC was created for that traffic.

It is built on the same technology as the DPU but trimmed down. An NVIDIA engineer quoted in the Japanese trade press explained that the full DPU had gained so many functions that it needed external power, whereas the SuperNIC is designed to run on the 75 watts a PCIe slot provides. It keeps the features that matter for AI networking and drops the rest.

The Register's description when it launched was that it is designed to work alongside NVIDIA's own Spectrum-4 switches as part of the Spectrum-X platform. That is the detail to remember.

Sources: BusinessNetwork.jp on the SuperNIC, The Register: making sense of NVIDIA's SuperNIC

The controversial part

The SuperNIC is as much a strategy as a product. Its best behaviour appears when it is paired with NVIDIA switches. Buy it and you have, in practice, also chosen your switch vendor. For NVIDIA that is the point: GPU, network card and switch from one supplier.

Competitors dispute that this is necessary. When the SuperNIC launched, a Broadcom executive told The Register that the same result could be achieved with Broadcom's switch chips and whichever DPU the customer prefers.

Both positions are commercial. The honest summary is that an all-NVIDIA network is the path of least integration effort, and an open one is the path of least dependence. We made the wider version of this argument in HPE Slingshot vs InfiniBand vs Spectrum-X.

Do you need them?

Your setup Our read
🖥️ One or two GPU servers, one team Standard NICs Good ordinary networking is enough
🗄️ A cluster of several 8-GPU servers training together SuperNICs The server-to-server fabric starts to matter
🏢 A shared platform serving several departments or customers Add DPUs Isolation and security become the reason
💾 GPUs waiting on storage Consider DPUs Storage traffic is where they are most often justified
🏭 A rack-scale system such as an NVL72 Already decided. They are part of the reference design.

Three questions for the quote

  1. Which lines on this quote are SuperNICs and DPUs, and what would the server lose without each?
  2. Do these cards require a particular brand of switch to deliver their benefit?
  3. Who configures and maintains the software that runs on the DPU?

The third question is the one that gets forgotten. A DPU is a computer. It needs firmware, updates and someone who understands it. A feature nobody has configured is an expensive network card.

Related on SourceIT

SourceIT Approved

SourceIT Approved

Researched and verified by the SourceIT team for Singapore buyers.

Share This Article
Facebook

More from our blog


← Older Post Newer Post →