News

In the ever-evolving landscape of artificial intelligence, the demand for more powerful, efficient, and scalable infrastructure has reached unprecedented heights. NVIDIA, a pioneer in accelerated computing, has answered this call with the launch of the BlueField-4 data processing unit (DPU), a groundbreaking innovation that is set to unleash the era of gigascale AI.

The BlueField-4 DPU is not just an incremental upgrade; it is a complete reimagining of what AI infrastructure can be. Powered by a combination of the NVIDIA Grace CPU and NVIDIA ConnectX-9 networking technology, it delivers a staggering 6x the compute power of its predecessor, the BlueField-3. This leap in performance enables AI factories to scale up to 4x larger than ever before, opening the door to trillion-token workloads and advanced agentic AI systems that were once thought impossible .

One of the most transformative aspects of the BlueField-4 is its role in enabling AI-native storage infrastructure. As AI models grow to trillions of parameters and engage in complex multi-step reasoning, they generate vast amounts of context data stored in key-value (KV) caches. Traditional storage solutions struggle to handle this data efficiently, leading to bottlenecks that hinder real-time inference. The BlueField-4 powers the NVIDIA Inference Context Memory Storage Platform, a purpose-built solution that extends AI agents' long-term memory and enables high-bandwidth sharing of context across clusters of rack-scale AI systems. This innovation boosts tokens per second by up to 5x and improves power efficiency by the same margin, eliminating GPU stalls and unlocking the full potential of large-context inference .

Security and multi-tenancy are also at the forefront of the BlueField-4's design. In today's multi-tenant AI environments, ensuring data isolation and protection is paramount. The BlueField-4 integrates advanced security features, including NVIDIA DOCA Argus, which provides AI runtime security without compromising performance. It supports secure bare-metal compute instances with stronger isolation, allowing enterprises to run sensitive workloads with confidence. Additionally, its multi-tenant networking capabilities enable efficient resource sharing across different users and applications, maximizing infrastructure utilization while maintaining strict security boundaries .