Nvidia Unveils Open-Source GPU Storage Tech Amid Industry Shift

Nvidia has released an open-source version of its cuFile storage stack, which it claims can improve data access times for graphics processing units (GPUs) by up to 95%. The company has partnered with Google, Intel, and Meta to make the technology available under a new GitHub organization called the Accelerated IO Special Interest Group.

cuFile allows a GPU to read from and write to storage directly, rather than relying on the CPU to fetch data. This can improve performance by reducing the time it takes for a GPU to access data. Nvidia claims that its technology can pull data from storage in microseconds, with a single server able to hit 230 million random reads per second at 512 bytes each.

The company has also launched an initiative called Storage-Next, which aims to create a new industry standard for GPU storage. More than 40 companies, including DDN, Kioxia, and Micron, have joined the effort to develop a massively parallel processor that can issue its own requests.