NVIDIA BlueField-3 DPU

NVIDIA BlueField-3 DPU: Transforming Infrastructure at 400Gb/s

A New Era of Infrastructure Acceleration

The NVIDIA BlueField-3 DPU marks a new chapter in data center evolution.

As the third-generation infrastructure compute platform from NVIDIA, it empowers organizations to build software-defined, hardware-accelerated IT environments across cloud, edge, and core infrastructure.

With network speeds up to 400 gigabits per second (Gb/s) over Ethernet or InfiniBand, BlueField-3 delivers exceptional performance for modern, distributed computing.

Performance Redefined: High-Speed Networking & Compute

BlueField-3 integrates powerful components that radically improve compute and data throughput:

• Up to 400Gb/s connectivity via Ethernet or NDR InfiniBand
• 16 Armv8.2+ A78 Hercules CPU cores with 8MB L2 and 16MB LLC
• 256-thread programmable data path acceleration for I/O-intensive tasks
• Onboard 32GB DDR5 memory and 128GB SSD storage
• PCIe Gen 5.0 x32 lanes, supporting both self-hosting and server-hosting modes

These specifications make BlueField-3 ideal for latency-sensitive, data-intensive, and multi-tenant workloads.

BlueField-3 delivers unprecedented levels of bandwidth and compute density in a compact and efficient design.

With dual network ports capable of handling up to 400Gb/s and a PCIe Gen5 x32 host interface, it eliminates network bottlenecks in high-throughput environments.

The powerful 16-core Armv8.2+ A78 CPU is paired with a 256-thread programmable accelerator optimized for concurrent processing, enabling it to manage millions of packets per second without impacting the host.

This architecture is ideal for latency-sensitive applications, such as real-time AI inference, distributed databases, and parallel computing clusters, ensuring seamless data movement and minimal I/O overhead.

Smart Offloading: Freeing CPU Resources for Business Applications

One of BlueField-3’s core strengths is its ability to offload and accelerate infrastructure tasks. These include:

• Software-defined networking (SDN): Offloads overlay networks, NAT, load balancing, and cloud network functions
• Storage acceleration: Optimized for NVMe™, NVMe/TCP™, and NVMe over Fabrics (NVMe-oF™)
• Security functions: Processes IPsec, TLS, MACSec, and firewall logic independently
• System management: Handles telemetry, orchestration, and configuration without burdening the host CPU

This separation between infrastructure logic and application execution enables improved performance and lower CPU utilization across the data center.

Traditionally, CPUs are overwhelmed by managing networking, storage, and security layers.

BlueField-3 changes that by offloading these infrastructure tasks onto the DPU, freeing up CPU cycles for application-level processing.

This includes handling tasks like traffic classification, overlay tunneling, storage virtualization, encryption, and telemetry collection.

As a result, organizations can achieve lower latency, higher throughput, and more deterministic performance.

For businesses operating at scale whether in the cloud, enterprise, or telecom sectors this offloading model directly translates to cost efficiency, scalability, and improved service-level agreement (SLA) compliance.

Uncompromising Security: Built for Zero-Trust Architecture

BlueField-3 implements a full suite of hardware-based security mechanisms, critical for today’s distributed environments:

• Secure boot with hardware Root of Trust (RoT)
• Firmware protection and flash encryption
• Support for AES-GCM 128/256-bit, AES-XTS 256/512-bit, and public key encryption (PKA)
• Functional isolation for tenant workloads in multi-tenant cloud environments
• True random number generator (TRNG) and device attestation

With these features, organizations can adopt zero-trust models, ensure data-in-motion and data-at-rest protection, and maintain compliance and integrity across their digital infrastructure.

Cybersecurity is a growing concern across all industries.

BlueField-3 is designed with zero-trust security in mind, providing multiple layers of protection through hardware-enforced features.

Secure boot mechanisms verify firmware integrity at startup, while hardware root-of-trust ensures that only signed code is executed.

Full-stack encryption capabilities protect data both in transit and at rest, supporting AES-GCM and AES-XTS standards.

It also offers device attestation, true random number generation, and microsegmentation.

These features make BlueField-3 ideal for highly regulated industries like finance, healthcare, and government, where compliance and data integrity are non-negotiable.

AI and HPC Ready: Built for the Workloads of Tomorrow

Designed to meet the compute and network demands of AI supercomputing, hyperscale clouds, and HPC workloads, BlueField-3 includes:

• GPUDirect® and GPUDirect Storage (GDS) for direct GPU communication
• MPI Tag Matching and All-to-All communication engines for parallel processing
• Ultra-efficient RoCE and Zero Touch RoCE
• BlueField-3 SuperNIC variant optimized for high-throughput, low-latency GPU-to-GPU communication

This makes BlueField-3 a perfect match for next-gen AI inference, training clusters, and large-scale scientific simulations.

Modern AI and HPC workloads demand more than raw power they require network-aware compute with minimal latency and high determinism.

BlueField-3 meets these needs with advanced features like GPUDirect and GPUDirect Storage, enabling direct memory access between GPUs across servers.

The inclusion of All-to-All engines and MPI tag matching enhances inter-GPU communication for distributed training or simulations.

These capabilities make it perfect for hyperscale AI clusters, scientific computing, and high-frequency trading environments, where data movement is as critical as computation.

BlueField-3 ensures that data pipelines are fast, secure, and optimized for AI acceleration.

Fully Programmable with NVIDIA DOCA™

The NVIDIA DOCA™ software framework provides a unified and secure environment for developers to:

• Build custom acceleration pipelines
• Control, monitor, and manage networking and storage behavior
• Ensure backward compatibility across BlueField generations
• Extend hardware capabilities using DOCA SDK, APIs, and microservices

With DOCA, BlueField-3 becomes more than just hardware—it becomes a platform for infrastructure innovation.

Flexibility is critical in today’s rapidly shifting digital landscape.

The NVIDIA DOCA™ framework enables developers to write and deploy applications that directly utilize BlueField’s acceleration capabilities.

Whether for networking, storage, security, or telemetry, DOCA allows for custom, high-performance data paths that match the unique demands of any organization.

With an extensive SDK, APIs, container support, and backward compatibility with previous BlueField generations, developers can future-proof their infrastructure while maintaining operational consistency.

DOCA also enables rapid deployment of third-party services, such as intrusion detection or traffic monitoring, making it a comprehensive programmable platform.

Sustainable and Scalable Data Centers

By offloading and consolidating infrastructure tasks into the DPU, BlueField-3 enables:

• Lower total cost of ownership (TCO)
• Reduced power consumption
• Better resource utilization
• Seamless scaling for modern cloud and edge computing models

It not only enhances performance, but also promotes data center sustainability, making it future-ready.

Real-World Use Cases

Domain	Applications
Cloud Networking	SDN acceleration, overlay networks, NAT, load balancing
Storage	NVMe/TCP, NVMe-oF™, elastic block storage, HCI
Security	Firewall, microsegmentation, DDOS prevention, platform security
AI & HPC	Multi-tenancy, AI job isolation, GPUDirect
Telco & Edge	vRAN, edge gateways, virtual network functions (VNF), microservers

 

BlueField-3 is versatile and applicable across a wide range of industries and workloads.

In cloud networking, it accelerates virtual switching, SDN, and service chaining.

In storage, it enhances performance of NVMe-oF™ and hyper-converged infrastructure platforms. For cybersecurity, it provides inline traffic inspection, encryption, and microsegmentation.

In AI and HPC, it improves job isolation, data movement, and GPU communication.

In telecom, it supports virtual RAN, edge microservers, and network slicing.

This wide applicability makes BlueField-3 a foundational component for any modern, distributed computing ecosystem that requires performance, security, and scalability.

Summary: Why Choose BlueField-3?

• End-to-end infrastructure acceleration
• AI, cloud, and HPC optimized
• Industry-leading security framework
• Unparalleled network throughput and I/O efficiency
• Programmable with DOCA and future-proof by design