Nvidia's cuFile Technology: A Game-Changer for AI Workloads

Nvidia has been working on a technology called cuFile, which aims to improve performance in AI workloads by allowing GPUs to access storage directly. The company announced that it will be open-sourcing this technology and making it available within its CUDA ecosystem.

CuFile is not a driver patch for GeForce RTX graphics cards that would turn an SSD into VRAM, as some had speculated. Instead, it allows the GPU to initiate and control storage operations on its own, which can help reduce latency in AI applications.

The technology was showcased at GTC 2026 earlier this year, where dedicated GPU threads could initiate I/O and access datasets that do not fit directly into local memory. This can be beneficial for tasks such as training on large datasets or running inference of models like Kimi K3.