Acronym · Compute & AI Hardware
DPU (Data Processing Unit)
A programmable network-attached processor that offloads networking, storage and security work from the host CPU, typically combining Arm cores, a NIC pipeline and hardware accelerators on one device.
- Also known as
- SmartNIC, IPU (Intel usage)
- Type
- Acronym
- Domain
- Compute & AI Hardware, Processor classes
Notes
DPUs matter in AI clusters because east-west traffic and storage access would otherwise steal host cycles from the accelerators.
Overview
A DPU, or data processing unit, is a programmable system on chip that sits at a server's network connection and takes over the networking, storage and security work that would otherwise consume the host CPU's cores. Cloud and AI data centres use DPUs so that infrastructure software runs on its own processor, isolated from customer workloads. The device is usually sold as a network adapter card, and vendors give similar parts other names, such as Intel's infrastructure processing unit.
How it works
A DPU combines three things on one chip: a high-speed network interface, a cluster of general-purpose CPU cores (commonly Arm), and fixed-function or programmable engines for packet processing, encryption, compression and storage protocols. Packets arriving from the network are parsed and steered in hardware; the cores run a full operating system and the control software for virtual switching, firewalls, telemetry and remote storage.
In a typical cloud server the DPU presents virtual network and storage devices to the host over PCI Express, while the real work of encapsulating traffic, enforcing policy or reaching remote drives over the network happens on the card. Because the DPU has its own processor, root of trust and secure boot, the operator's infrastructure software stays isolated from whatever runs on the host, and every host core can be rented to tenants.
Vendors take different routes to programmability. NVIDIA's BlueField line pairs Arm cores with its own network hardware and is programmed through a software framework called DOCA; AMD Pensando parts use packet pipelines programmed in the P4 language; Intel's IPU E2100 uses Arm Neoverse cores. The trade-off is cost, power and a separate software stack to maintain, which pays off mainly at large scale.
Examples
- An NVIDIA BlueField-3 DPU running software-defined networking, storage and security services at 400 Gb/s line rate in an AI server.
- An AMD Pensando DPU with a P4-programmable pipeline used for front-end networking in a hyperscale data centre.
- Intel's IPU Adapter E2100, which uses 16 Arm Neoverse N1 cores to offload infrastructure tasks and accelerate NVMe storage transport.
- A cloud operator presenting remote block storage to a tenant's virtual machine as if it were a local drive, with the DPU handling the network storage protocol.
Why it matters
Hyperscale cloud providers, AI cluster builders and storage system makers adopt DPUs to win back host CPU capacity and to separate the infrastructure from the tenant. NIC and DPU architects, dataplane software engineers and cluster infrastructure teams work with them.
In the chip lifecycle, the DPU's offload engines and core count are set at architecture, its packet pipeline is a major verification effort, and after production it is validated in full server and cluster systems where software and firmware updates continue for years.
It matters most to: AI Cluster Infrastructure Engineer, DPDK and Dataplane Software Engineer, Network Switch and NIC Silicon Verification Engineer, NIC and DPU Architect, Server SoC Architect
Common confusions
- Not the same as a SmartNIC
- A SmartNIC is a network card with some offload capability; the term is broader and older. A DPU is the processor at the heart of the more capable SmartNICs, with general-purpose cores able to run a full operating system.
- Not the same as an IPU
- Intel calls its comparable device an infrastructure processing unit. The idea is close to a DPU, though the same three letters are also used by other companies for unrelated AI processors.
- Not the same as a GPU
- A GPU runs the application's parallel compute, such as AI training. A DPU moves and protects the data around it and does not run the tenant's workload.
Sources
- What's a DPU? (Kevin Deierling, 2020) (opens in a new tab), NVIDIA, 2026-10-22
- BlueField Networking Platform (opens in a new tab), NVIDIA, 2026-10-22
- AMD Pensando DPU Technology (opens in a new tab), AMD, 2026-10-22
- Intel Infrastructure Processing Unit (Intel IPU) (opens in a new tab), Intel, 2026-10-22
Related terms
Part of the Compute & AI Hardware glossary, with every term in the topic and the lifecycles, tools and standards that use them.
Where it appears on the Hub
Used on 6 pages of the Hub. In each group, the pages that mention it most come first.
Roles
Skills
Suggest an improvement
Suggest an improvement to this entry
A clearer definition, an example, a correction or a source an editor can check. Editors review every suggestion before the entry changes.
Last reviewed . Written by The Semiconductor Hub editorial team. The full glossary