Building the "Most Powerful Brain" for intelligent agent AI: NVIDIA launches the BlueField-4 STX storage architecture

K-LinePoet · 2026-03-24T07:28:06+00:00

NVIDIA unveiled the BlueField-4 STX modular storage architecture at its 2026 GTC conference to address performance bottlenecks in traditional storage for AI agent inference. The architecture improves processing efficiency by positioning storage nodes closer to compute resources, supporting higher throughput and faster response times while significantly enhancing GPU utilization. Multiple leading storage vendors and cloud service providers will adopt this technology, with plans to launch STX platform-based systems in the second half of the year.

K-LinePoet

2026-03-24 07:28:06

Abstract generation in progress

IT Home March 17 News, today (March 17), at the 2026 GTC Conference held in San Jose, California, NVIDIA announced the BlueField-4 STX modular storage reference architecture for intelligent agents AI (Agentic AI) long-context reasoning.

IT Home quotes a blog post stating that traditional data centers offer large-capacity general storage but lack rapid response capabilities, making it impossible to support seamless cross-step and cross-tool interactions for AI agents. As contextual information increases, traditional storage paths significantly slow down AI inference speed and greatly reduce GPU utilization.

The STX architecture addresses this pain point by bringing storage nodes closer to computing resources and enabling scalable access, facilitating higher throughput and response speeds in inference, training, and analysis phases of AI agent factories.

Its first rack-level deployment solution integrates the new NVIDIA CMX in-memory context storage platform. This platform extends GPU memory through high-performance context layers, increasing the number of tokens processed per second by up to 5 times compared to traditional general storage devices.

The STX architecture is built on NVIDIA’s Vera Rubin platform, featuring the BlueField-4 processor optimized for storage (internally integrated with Vera CPU and ConnectX-9 super network card), and works in conjunction with Spectrum-X Ethernet, DOCA, and AI Enterprise software.

In terms of actual performance, this architecture offers four times the energy efficiency of traditional CPU architectures and doubles data ingestion speed (pages processed per second) when handling enterprise-level AI data.

NVIDIA founder and CEO Jensen Huang emphasized that to reshape computing infrastructure to match the evolution of intelligent agents AI, new storage forms are needed to handle massive contexts and support continuous learning.

Top storage providers like Dell, HPE, and IBM are currently co-designing the next-generation AI infrastructure based on STX. Additionally, leading cloud service providers such as Oracle Cloud (OCI) and Mistral AI plan to adopt this technology first. Related partners will officially deliver systems based on the STX platform to the market in the second half of this year.

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.