Nvidia BlueField
Nvidia data processing units with their own CPU cores and memory
Nvidia BlueField is a line of data processing units (DPUs) that take network, storage and security work off the host CPU. The design began at Mellanox Technologies and passed to Nvidia when it bought Mellanox in March 2019. BlueField-2 launched at GTC 2020, with BlueField-3 and BlueField-4 announced at GTC 2021.
- Nvidia bought Mellanox Technologies in March 2019 for US$6.9 billion
- BlueField-2 launched at GTC 2020 after VMworld 2019 samples
- Cards carry their own CPU cores and run a separate operating system
- Can switch PCIe into host mode to control a GPU over the network
- BlueField-3 added HBM and DDR5 memory support
What it is
Nvidia BlueField is a family of data processing units, or DPUs: add-in cards that carry their own processor cores and memory so they can run datacentre jobs without leaning on the host CPU. Mellanox Technologies created the design, and Nvidia took it over when it bought Mellanox in March 2019 for US$6.9 billion. The first cards shipped under the Nvidia name were BlueField-2, shown at VMworld 2019 and formally launched at GTC 2020. They are aimed at datacentres and high performance computing, where latency and bandwidth decide how well work runs.
How it came about
The BlueField intellectual property started life inside Mellanox Technologies. Nvidia acquired Mellanox in March 2019 for US$6.9 billion, bringing the design with it. BlueField-2 samples went out for review soon after the VMworld 2019 announcement, and the card was officially launched at GTC 2020. That same event saw the BlueField-2X, which pairs the DPU with an Ampere generation GPU on one card. At GTC 2021 Nvidia announced BlueField-3 and BlueField-4, giving tentative launch years of 2022 and 2024.
Inside the chip
A BlueField card is more than a network interface controller. It carries CPU cores, usually ARM or MIPS based, plus its own memory, so it can take over tasks the host processor would otherwise handle. It runs an operating system entirely separate from the host, which cuts software overhead and lets each card work on its own. That independence also enables remote control of machines that lack such features. The card can even switch its PCIe bus into host mode, linking through a PCIe bridge to another card, such as a compute accelerator, so a GPU can be managed over the network at high bandwidth.
Where it went
BlueField cards are built for datacentres and high performance computing, where low latency and high bandwidth matter. Because each card runs its own system, it can handle network, storage and security work while the host CPU is left free. The X variants target large GPU clusters, keeping traffic off the PCIe bus and off the CPU so compute stays available for other work. The extra connectivity lets a dataset be shared across several nodes when it is too big for one machine's memory.
Versions and family
BlueField-2 was the first Nvidia-built card, followed by BlueField-2X, which adds an Ampere generation GPU on the same board. BlueField-3 and BlueField-4 were announced at GTC 2021 for 2022 and 2024. BlueField-3 brought support for HBM and DDR5 alongside the usual DDR4. The X cards place a 100 class Nvidia datacentre GPU on the same PCB as the DPU. The H100 CNX and A100 EGX look similar but are separate products: they use a generic ConnectX network interface controller rather than BlueField system on a chip integration.
Why it matters
BlueField moved functions that once sat on the CPU onto the card itself, changing how datacentre machines handle networking, storage and security. Its ability to run independently of the host, and to act as a PCIe host toward other cards, gave operators new ways to manage hardware remotely and to drive GPUs over the network. The line grew out of Mellanox work and became a Nvidia product family spanning several generations.
Questions people ask about Nvidia BlueField
What is a BlueField card?
It is a data processing unit, or DPU: an add-in card with its own CPU cores and memory that takes over networking, storage and security work normally done by the host processor. Nvidia designs the BlueField line, which began at Mellanox Technologies before Nvidia bought that company in March 2019.
How does it differ from a normal network card?
A BlueField card has CPU cores, usually ARM or MIPS based, and its own memory, and it runs an operating system separate from the host. It can offload CPU work, manage systems remotely, and even set its PCIe bus to host mode so it can drive another card such as a GPU over the network.
What is a BlueField-X card?
It is a hybrid board that puts a 100 class Nvidia datacentre GPU on the same PCB as the BlueField DPU. Keeping both on one card lets a GPU cluster communicate at high bandwidth without crossing the PCIe bus, so the host CPU is free for other work.
Are H100 CNX and A100 EGX BlueField cards?
No. They are NIC and GPU hybrid cards that look much like a BlueField-X product, but they are separate designs. Instead of BlueField system on a chip integration, they carry a generic ConnectX network interface controller.
What memory do these cards use?
BlueField cards typically use DDR4. With the release of BlueField-3, support widened to include HBM and DDR5 as well.
Good reading and sources
- Nvidia completes 'homerun deal' after closing $7 billion acquisition of MellanoxCNBC, Article
- Nvidia Reveals BlueField-3, BlueField-4 DPUs: 400-800 Gbps, 22-64B TransistorsTom's Hardware, Article
The facts on this page come from the Wikipedia article "Nvidia BlueField", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.