Nvidia Open-Sources cuFile API to Use SSDs as Extra GPU Memory
Nvidia has open-sourced its cuFile API and the underlying storage software stack for GPUDirect Storage (GDS) at the Black Hat 2026 conference. This technology enables direct data transfers between NVMe SSDs and GPU memory, bypassing the CPU and system RAM. By allowing SSDs to act as direct extensions of GPU memory, this development can significantly reduce latency and prevent "GPU starvation" in AI workloads. It also has the potential to reduce game stuttering by accelerating data access speeds for graphics rendering. The cuFile API utilizes Direct Memory Access (DMA) to achieve millisecond-level data retrieval directly from NVMe drives. Additionally, Nvidia launched the "Storage-Next" initiative, collaborating with over 40 major storage and flash memory vendors like Micron and Kioxia.
## BACKGROUND
Traditionally, data must be read from storage into a system memory buffer before being copied to the GPU, creating a bottleneck. GPUDirect Storage (GDS) solves this by establishing a direct path between storage and GPU memory, which minimizes CPU overhead and maximizes throughput.