GPU-driven storage has become a chokepoint for AI workloads, and NVIDIA used the Future of Memory and Storage show in Santa Clara this week to hand the industry an open toolkit for fixing it.
The centerpiece is cuFile, the software behind GPUDirect Storage that allows GPUs to move data without CPU involvement. NVIDIA placed the code in a public GitHub organization run jointly with Google, Intel, and Meta, and says access times drop into the microsecond range once systems adopt the stack.
It also rallied the storage world behind a new effort called Storage-Next. More than 40 companies signed on, among them DDN, KIOXIA, and Micron, with a charter to standardize how GPU-driven storage behaves across controllers, cooling, thermal design, and orchestration.
The pitch is that the old rules for deciding whether data lives in memory or on disk no longer hold. NVIDIA cites its own testing where the Vera CPU handled a compression and encryption pipeline at 3.21x the rate of an x86 processor, meaning storage gear can soak up AI traffic without adding racks of compute.
The wider portfolio includes SCADA, a framework for letting GPU clusters request exactly the data they need, and the Vera BlueField-4 STX storage processor. NVIDIA calls the whole effort extreme codesign, with storage now as central to AI progress as the silicon that trains the models.
