NVIDIA unveiled the NVIDIA BlueField-4 data processing unit (DPU) at the GTC conference in Washington, D.C. This new DPU will improve AI data storage, network transmission, and security through software-defined acceleration, transforming data centers into secure, intelligent AI infrastructure. NVIDIA describes it as the core engine driving an "AI factory operating system."
The launch of BlueField-4 is in response to the ever-increasing scale of AI factories and the explosive demand for processing trillion-token workloads.
Core architecture: Grace CPU + ConnectX-9, significantly improving performance and scale
BlueField-4 has undergone a major upgrade in hardware architecture, integrating:
• NVIDIA Grace CPU: Provides powerful general computing capabilities.
• NVIDIA ConnectX-9 Networking Technology: Achieve data throughput of up to 800Gb/s.
NVIDIA claims that the computing power of the newly launched BlueField-4 is 6 times higher than that of the BlueField-3, and it can support AI factory infrastructure that is 4 times larger in scale.
BlueField-4 also features multi-tenant networking, fast data access, AI runtime security, and cloud elasticity, and natively supports NVIDIA DOCA microservices, a containerized service framework designed to secure, scale, and simplify AI deployment and operations.
The BlueField-4 platform supports a multiservice architecture and native service function chaining, seamlessly integrating and managing multiple network, security, and storage services within a single framework.
Security: ASTRA architecture enables zero-trust tenant isolation
Security is another key focus of BlueField-4. Based on the NVIDIA BlueField Advanced Secure Trusted Resource Architecture (ASTRA), it enables service providers to create secure bare-metal compute instances with zero-trust tenant isolation while maintaining full control over their software-defined infrastructure.
NVIDIA emphasizes that the BlueField platform can provide consistent performance, efficiency, and security across its entire AI infrastructure product portfolio, ranging from RTX PRO servers, HGX, DGX, GB200/GB300 systems to data center-level designs such as Enterprise AI Factory, AI Data Platform, and DGX SuperPOD.
ConnectX-9 SuperNIC: 800Gb/s network card for Spectrum-X
In conjunction with BlueField-4, NVIDIA also launched the NVIDIA ConnectX-9 SuperNICs intelligent network interface card. Designed specifically for the NVIDIA Spectrum-X Ethernet platform, this card provides ultra-low-latency 800Gb/s network connectivity, designed to maximize the efficiency of gigascale AI infrastructure and enhance RoCE (RDMA over Converged Ethernet) performance, delivering consistent and predictable network performance for the most demanding AI workloads.
Widely adopted by the ecosystem, DOCA ensures backward compatibility
NVIDIA stated that the BlueField platform has achieved broad ecosystem adoption, with numerous partners planning to integrate BlueField-4 to build the next-generation accelerated, secure, and software-defined data center platform. Through the DOCA framework, applications currently accelerated on the BlueField platform will run seamlessly on BlueField-4, achieving significant performance improvements.
• Server and storage facilities: Cisco, DDN, Dell, HPE, IBM, Lenovo, Supermicro, VAST Data, WEKA, etc.
• Network Security Factory: Armis, Check Point, Cisco, F5, Forescout, Palo Alto Networks, Trend Micro, etc. (leveraging BlueField to provide zero-trust, AI-powered execution-time security).
• Cloud and AI providers: Akamai, CoreWeave, Crusoe, Lambda, Oracle Cloud Infrastructure, Together.ai, xAI, etc. (leveraging DOCA microservices to accelerate multi-tenant networking, data movement, and security).
• Cloud native and basic software platform: Canonical, Mirantis, Nutanix, Rafay, Red Hat, Spectro Cloud, SUSE and more.
• Global System Integrators: Accenture, Deloitte, World Wide Technology, etc.
Timeline for listing
NVIDIA BlueField-4 is expected to be available in early 2026 as part of the NVIDIA Vera Rubin platform.
