NVIDIA Introduces BlueField-4 STX for AI Storage and Context Data

NVIDIA targets storage constraints in AI inference

NVIDIA has introduced BlueField-4 STX, a storage reference architecture designed for AI workloads that must continuously move, retrieve and manage large volumes of context data.

In a technical blog post, NVIDIA said agentic AI and long-context inference place increasing demands on storage infrastructure. Tasks including KV-cache management, context-memory handling, compression, encryption, integrity checks and data recovery can consume substantial CPU capacity, according to the company. NVIDIA’s argument is that shifting more of this work closer to the storage data path can preserve compute resources for AI services.

BlueField-4 STX is built around NVIDIA’s Vera CPU and ConnectX-9 SuperNIC. NVIDIA describes it as an AI-native storage processor and a modular reference design for storage services that support inference context and cache data.

NVIDIA reports gains in storage-oriented benchmarks

NVIDIA’s technical blog reported microbenchmark results for the Vera CPU in the Vera BlueField-4 STX Storage Processor. The company said the processor achieved up to 3.21 times higher throughput for compression and encryption than the platform used as its comparison system.

The company also cited performance improvements in integrity-checking and data-recovery operations. NVIDIA presents these results as evidence that the architecture can provide more CPU headroom to storage software used in AI environments.

Those figures are vendor-published results rather than independent measurements. Actual performance will depend on the storage software, system configuration, data characteristics and deployment design, and NVIDIA’s post does not establish that identical gains will apply across all implementations.

Broader claims include token throughput and ingestion speed

In its newsroom announcement, NVIDIA said BlueField-4 STX combines Vera CPU technology with the ConnectX-9 SuperNIC, alongside support for Spectrum-X Ethernet and NVIDIA’s DOCA software framework.

NVIDIA claimed the architecture can provide up to five times token throughput, up to four times energy efficiency and twice the data-ingestion speed in relevant AI-storage scenarios. The company attributed those outcomes to its stated configurations and workloads, so the numbers should be interpreted as product claims rather than general performance expectations.

Tom’s Hardware reported that the architecture addresses a data-access bottleneck in agentic inference. In multi-step AI tasks, models and connected tools may repeatedly retrieve, update and use external information, making data movement and storage responsiveness more significant to overall service performance. The publication also described BlueField-4 STX as combining Vera CPU, ConnectX-9, Spectrum-X Ethernet and DOCA software.

Extending AI infrastructure into storage processing

BlueField-4 STX reflects NVIDIA’s effort to position storage processing as a more integrated part of AI infrastructure, rather than solely a back-end function. The design is intended to bring context handling, data movement and storage services closer to the needs of inference applications.

Its practical impact will depend on adoption by storage-software providers and system builders, as well as performance evidence from production deployments. For now, NVIDIA’s announcement makes storage and context management a more explicit element of its architecture for agentic and long-context AI.