Overview of NVIDIA BlueField DPU Features and Architecture

The article, based on the "Future Network: SmartNIC DPU Technology Whitepaper," introduces DPU as a data‑centric, I/O‑intensive specialized processor that integrates and surpasses traditional SmartNIC capabilities.

(1) Interface and Drivers Each BlueField DPU network port exposes two physical PCIe functions—one for the embedded Arm subsystem and one for the Host—requiring the mlx5 driver stack to be loaded on both Arm and Host servers, with OS detection on each host.

(2) Work Modes DPU can operate in Embedded CPU Function (ECPF) ownership, Restricted mode (an extension of ECPF), NIC mode (behaving like a standard NIC from the host perspective), and Separation mode, which allocates network functions between the Arm kernel and the x86 host kernel.

(3) Kernel Representation Model BlueField uses netdev representors to map physical and virtual functions, acting as tunnels or channels to connect virtual switches (e.g., OVS) on the Arm kernel with corresponding PF/VF on the host.

(4) Multi‑Host In multi‑host mode, each host interface can be split into up to four independent PCIe interfaces, all sharing a single physical port managed by a multi‑physical function switch (MPFS), with each host controlling its own e‑switch.

(5) Virtual Switch on DPU NVIDIA BlueField supports ASAP technology and includes an OVS package with ASAP2, enabling the Arm‑based virtual switch to offload data‑path processing and maintain standard OVS control plane operations.

(6) Network Bonding The DPU can configure network bonding transparently to the host, presenting a single PF while aggregating multiple interfaces for redundancy and increased throughput.

(7) Upstream MTU Configuration To set port MTU in SmartNIC mode, the host side of the BlueField DPU must be limited, after which the MTU can be adjusted via the upstream link representor (p0 or p1).

(8) Link Aggregation Similar to network bonding, link aggregation combines multiple interfaces into one, providing fault tolerance and bandwidth aggregation while exposing only a single PF to the host.

(9) Scalable Functions (SF) SFs are lightweight functions with dedicated resources (queues, CQ, EQ) that do not share or preempt parent PCIe function resources, coexisting with SR‑IOV virtual functions without BIOS support.

(10) RDMA Stack Support The Arm Linux system includes a full RDMA stack, and BlueField DPU supports RDMA (RoCE or InfiniBand) in both Separate Host Mode and Embedded CPU Mode.

(11) Host PF and VF Parameter Control Users can set MAC addresses, link states (Up/Down/Follow), query configurations, and disable host PF exposure for storage‑only or VirtIO use cases, noting that disabling PF also disables SR‑IOV and related SF usage.

(12) DPDK Support BlueField DPU supports the Data Plane Development Kit (DPDK), providing a fast packet‑processing development platform and APIs.

(13) SNAP NVIDIA BlueField SNAP enables hardware‑accelerated NVMe storage virtualization, presenting network storage as local NVMe SSDs and integrating with standard NVMe drivers on the host.

(14) Regular Expression Acceleration BlueField‑2 DPU offers high‑speed RegEx acceleration, allowing hosts to offload multiple RegEx jobs via DPDK APIs.

(15) Compression Acceleration The DPU provides high‑performance compression/decompression offload, operating in parallel with net, vDPA, and RegEx workloads.

(16) Public‑Key Acceleration Integrated PKA engines accelerate cryptographic operations (e.g., SSL, PK signatures) for security‑critical applications.

(17) IPsec Offload BlueField can transparently offload IPsec encryption/decryption from the host CPU, running IPsec in the DPU’s software layers.

(18) QoS Configuration The DPU supports rate limiting for Host PF/VF, SF, VF groups, and QoS groups, enabling fine‑grained traffic management.

(19) VirtIO‑net Emulated Devices The DPU can create up to 16 VirtIO‑net PCIe PF devices or 504 VF devices, allowing hot‑plug of virtual NICs in the host system.

(20) Deep Packet Inspection (DPI) DPI, part of the DOCA SDK, provides advanced packet filtering to detect threats, malware, policy violations, and supports load balancing.

(21) Shared RQ Mode To reduce memory consumption when scaling many representors, PF/VF/SF can share a common receive queue, saving up to several gigabytes of memory.