Nvidia Hands Over Control: Open-Sourcing cuFile and Shaping Industry Standards
Nvidia has open-sourced its cuFile APIs and storage software stack, marking a significant shift in how data reaches processors. The company's vice president of storage technology, Jason Hardy, stated that the result is data pulled from storage in microseconds.
The new framework, SCADA (scaled accelerated data access), allows GPUs to build their own storage requests and handle completions, keeping hundreds of thousands of operations in flight at once.
According to Micron, a single server hit 230 million random reads per second at 512 bytes each using three H100 GPUs. This is roughly 95 percent of what those drives are rated to deliver. The CPU sat essentially idle during this process.
Nvidia's stated reason for open-sourcing cuFile is security and Linux interoperability. Letting an application talk to a drive directly requires careful privilege separation, handled with standard Linux mechanisms.
The company has also launched Storage-Next, an initiative that spans more than 40 storage and flash makers, controller vendors, cooling specialists, and standards bodies.