[Digital Today reporter Chi-gyu Hwang (황치규)] Nvidia has released its cuFile API, which directly links GPUs and storage, as open source. On Aug. 4, Nvidia announced cuFile and an industry collaboration group called Storage-Next at the Future of Memory and Storage conference, SiliconANGLE reported.

cuFile is a technology that sends storage data directly to GPU memory without going through the central processing unit and system memory. The cuFile API supports secure access to storage data in milliseconds, the company said. cuFile was formally launched in July 2021 with CUDA Toolkit 11.4.

The technology uses direct memory access to move data from storage devices such as NVMe drives straight into GPU memory. In the process, the GPU can bypass the CPU and system memory, reducing memory access latency.

Nvidia said cuFile would be useful for faster processing of retrieval-augmented generation and agentic AI workloads.

Storage-Next includes 40 storage and flash memory companies, including DataDirect Networks, Kioxia and Micron. Storage-Next brings together storage makers, controller companies, thermal design and cooling firms, orchestration providers and standards bodies, with a focus on finding ways to optimise storage around GPUs.

Nvidia also presented SCADA as a foundational technology that allows massively parallel GPUs to fetch data with high bandwidth and very low latency.