Nvidia Open-Sources cuFile API to Accelerate GPU Storage Access
Nvidia has open-sourced its cuFile API, a key component of its GPUDirect Storage software. The move aims to accelerate GPU read/write capability for high-speed storage and enable millisecond data access.
The cuFile API enables direct, high-speed data access between local distributed storage and GPU memory. This allows GPUs to bypass the central processing unit and system main memory, reducing memory access delays. By using direct memory access (DMA), data can be moved directly from storage devices such as NVMe drives into GPU memory.
Nvidia's open-sourcing of cuFile is part of its larger industry initiative with Storage-Next, which includes 40 leading storage and flash memory vendors. The goal is to optimize memory and storage solutions for AI applications, which require massive data access at high bandwidth and extremely low delay.