NVIDIA pushes storage closer to GPUs for AI workloads

Rising AI workloads are pushing datasets and context windows beyond system memory, increasing pressure on storage systems to do more than add capacity. NVIDIA said AI agents and GPUs are creating thousands of concurrent storage operations, requiring systems to encrypt, compress, verify and reconstruct data without becoming bottlenecks.

NVIDIA cited benchmarks showing its Vera CPU, part of Vera BlueField-4 STX, delivers up to 3.21x higher throughput than an x86 CPU in a two-stage compression and encryption pipeline. The company framed accelerated storage as an active part of the AI data path, where decisions about keeping data in memory or on drives now play out in microseconds.

At the Future of Memory and Storage conference, NVIDIA announced it is open sourcing its cuFile APIs and the storage software stack beneath them, enabling GPUs to read from and write to storage directly. Google, Intel, NVIDIA and Meta are inaugural maintainers of the new open contribution site, with the APIs intended to support interoperability across software and hardware platforms.

NVIDIA is also working with storage vendors through Storage-Next, an initiative that includes over 40 leading storage and flash vendors, including DDN, KIOXIA and Micron. The effort includes SCADA, a scaled, accelerated data access framework designed to let GPUs pull only needed data from storage into high-speed memory while maintaining protected access through Linux-based security enforcement.