NVIDIA has introduced its modular reference architecture, BlueField-4 STX, aimed at enhancing data storage platforms for enterprises, cloud providers, and data center operators, specifically optimized for artificial intelligence tasks. Traditional data centers often rely on general-purpose storage solutions that, while capable of handling large volumes of data, struggle to deliver the rapid response times required by AI agents. These systems demand real-time data access and contextual memory, which the STX architecture intends to address by eliminating existing bottlenecks.
At the heart of the STX architecture is the NVIDIA BlueField-4 DPU, which integrates an NVIDIA Grace/Vera Arm processor, 128 GB of LPDDR5, a 512 GB SSD, and a NVIDIA ConnectX-9 SuperNic network adapter with a throughput of 1.6 Tbps, alongside a PCIe 6.0 switch featuring 48 lanes. Utilizing NVIDIA's DOCA microservices and NVIDIA AI Enterprise software, the STX architecture reportedly offers four times the energy efficiency compared to traditional CPU-based storage architectures. This innovative solution aims to facilitate a versatile data processing mechanism that accelerates the entire AI lifecycle, from training and analytics to inference powered by agents.
The first large-scale implementation of STX is the new NVIDIA CMX storage platform, which includes contextual memory to extend GPU memory capabilities. Major companies such as Cloudian, DDN, Dell Technologies, and IBM have announced support for NVIDIA STX. System production will be undertaken by AIC, Supermicro, and Quanta Cloud Technology (QCT), while various organizations, including CoreWeave and Lambda, plan to implement the platform. Solutions based on STX are expected to be available in the second half of this year.
This development positions NVIDIA as a leader in the AI infrastructure market, potentially reshaping competition among tech giants as they vie for dominance in high-performance data storage solutions tailored for AI applications.
Informational material. 18+.
