NVIDIA BlueField DPU Series: Architecture, Features, and Ecosystem Overview

Modern ultra‑large‑scale cloud technologies are driving data‑center architectures toward specialized processors that offload and accelerate virtualization, networking, storage, security, and other cloud‑native AI services; NVIDIA's BlueField DPU series embodies this trend.

BlueField‑2 is the current generation, while BlueField‑3 introduces a 400 Gb/s AI‑optimized DPU capable of delivering the performance of up to 300 CPU cores, supporting multi‑tenant cloud environments with hardware‑accelerated networking, storage, security, and management. BlueField‑4 is planned to integrate NVIDIA GPU functions and reach 800 Gb/s.

The DPU architecture merges a programmable data‑path, hardware accelerators for encryption/compression, and an ARM control processor. BlueField‑3 features 16 cores handling 256 threads, with key technical highlights in four domains:

1. Network Business – Enhanced RDMA, connection tracking, ASAP, and precise clock synchronization; RDMA provides zero‑copy, kernel bypass, no CPU involvement, message‑based transactions, and scatter/gather support.

2. Security Business – Full‑line‑rate 400 Gb/s encryption/decryption at IP, transport, and MAC layers; supports IPSec (AH, ESP, IKE) and TLS with speeds far exceeding CPU‑based implementations, freeing CPU cycles.

3. Storage Business – Supports block, file, object, and NVMe storage, with hardware‑offloaded encryption (e.g., AES‑XTS) and SNAP technology that virtualizes remote NVMe as local storage, delivering up to 18 M IOPS read/write and 80 Mpps virtual I/O acceleration.

4. Development Ecosystem – The DOCA SDK provides libraries, APIs, and tools for building software‑defined networking, storage, and security applications on BlueField DPUs, abstracting DPU programming similarly to CUDA for GPUs.

Overall, DPUs act as a third compute pillar alongside CPUs and GPUs, handling data‑plane and control‑plane tasks such as networking, storage, virtualization, and security, thereby alleviating CPU bottlenecks in modern data‑center workloads.