NVIDIA's BlueField-4 DPU to Deliver 800 Gbps AI Inference Processing for Vera Rubin Architecture

NVIDIA has unveiled its next-generation BlueField-4 Data Processing Unit (DPU), a specialized chip designed to accelerate data center infrastructure and AI workloads. Announced at the GTC event in Washington, the BlueField-4 marks a significant leap by doubling the AI inference processing throughput to 800 gigabits per second (Gbps), up from 400 Gbps in the current BlueField-3 model. This massive bandwidth increase is critical for handling the immense data flows required by modern AI models and large-scale simulations.

The new DPU is architected as part of a powerful compute node, pairing NVIDIA's own Grace CPU with an 800 Gbps ConnectX-9 SuperNIC (Smart Network Interface Card). NVIDIA claims this combination delivers a sixfold boost in overall compute processing capacity. This performance is targeted at enabling "AI factories"—large-scale, dedicated AI training and inference clusters—that can be up to four times larger than what's possible with previous technology. A flagship application for this architecture will be the Vera Rubin Observatory, which requires processing enormous datasets from its sky survey.

Scheduled for release in 2026, the BlueField-4 represents NVIDIA's continued push to offload and accelerate networking, storage, and security tasks from central CPUs. By handling these functions on the DPU, the main servers (CPUs and GPUs) are freed to focus exclusively on computation, dramatically improving overall system efficiency and scalability for the next wave of AI infrastructure.