Specifications
Product
NVIDIA BlueField-4 DPU
Network Throughput
800 Gb/s total (typically 2 × 400 Gb/s bi-directional)
Compute
64-core NVIDIA Grace CPU, Arm Neoverse V2 microarchitecture
Transistor Count
64 billion
Compute Uplift
6× the compute power of BlueField-3
Networking
Integrated NVIDIA ConnectX-9 SuperNIC
Host Interface
PCIe Gen 6 x16
Front-End
Ethernet (north-south traffic: storage, management, inference I/O)
East-West
ConnectX-9 SuperNIC pairing for 1.6 Tb/s GPU-to-GPU (Rubin generation)
Software
NVIDIA DOCA microservices framework
Offloads
Networking, data storage, cybersecurity, encryption, policy enforcement
Security
Line-rate hardware crypto acceleration, multi-tenant isolation, secure boot
Storage Processor
BlueField-4 STX storage processor variant for AI storage front-ends
Ecosystem
WEKA NeuralMesh, NVIDIA Magnum IO, DPU-accelerated storage clients
Platform
Vera Rubin AI factory rack scale-out
Form Factor
FHHL PCIe add-in card / OCP-adjacent designs per vendor
Power
Per-vendor; DPU offload reduces host CPU power for the same network service set
Use Case
AI factory operating system — networking, storage, security planes
Availability
Sampling / early production — quote for allocation
Overview
The NVIDIA BlueField-4 DPU is the infrastructure processor for the Rubin generation: an 800 Gb/s network endpoint with a 64-core Grace CPU that runs the software-defined networking, storage and security planes directly on the card. It is announced as part of the Vera Rubin platform, where ConnectX-9 handles east-west GPU traffic and BlueField-4 handles north-south infrastructure traffic.
The architectural shift is that BlueField-4 stops being a NIC-with-cores and becomes a server in its own right. With 64 Grace cores, 64 billion transistors and a PCIe Gen 6 host link, it has the capacity to terminate storage protocols, run encryption and policy in line, and host the micro-services that would otherwise consume host CPU cycles in every node of a rack.
That is what makes it relevant to storage vendors and service providers: WEKA's NeuralMesh, for example, pushes replication and data-path optimisation onto the DPU so they execute where packets flow. QS Compute supplies BlueField-4 DPUs alongside BlueField-3 for retrofit, the ConnectX-9 SuperNIC, and the switches these cards attach to.
Key Benefits
6× BlueField-3 compute in the same infrastructure slot means more network services per rack without touching host CPUs. 800 Gb/s matches the Rubin-generation north-south requirement so storage and management traffic stop competing with GPU traffic. DOCA microservices let you run a full storage or security service stack on the DPU, and line-rate crypto keeps multi-tenant AI factories compliant without a performance penalty.
Applications
AI factory storage front-ends and NVMe-oF target offload; multi-tenant service-provider networking with per-tenant isolation; in-line encryption and policy enforcement for sovereign AI clouds; and software-defined network functions formerly run on host CPUs.
Request a Quote — NVIDIA BLUEFIELD-4 DPU — 800 GB/S INFRASTRUCTURE PROCESSOR FOR AI FACTORIES
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →