your GPU has been waiting on your CPU this whole time — nvidia just gave it a shortcut

cuFile lets a GPU pull data straight off an SSD in microseconds, skipping the CPU hop that's been quietly starving AI training jobs of work to do.

Aliteq
Lena Fischer · AI & Local Compute Editor

The short version

Nvidia open-sourced cuFile and its storage software stack on August 4, 2026, at the FMS conference

The short version

cuFile uses direct memory access (DMA) to move data from NVMe drives straight into GPU memory, skipping the CPU and system RAM

The short version

It targets 'GPU starvation' — GPUs sitting idle waiting for data, especially bad for mixture-of-experts and RAG workloads

The short version

Google, Intel and Meta joined Nvidia as founding maintainers of the new Accelerated IO SIG on GitHub

The short version

A parallel 40+ vendor consortium called Storage-Next, including DDN, Kioxia and Micron, is aligning the industry around it

Aliteq

Read the full story

your GPU has been waiting on your CPU this whole time — nvidia just gave it a shortcut

 Read the full story on Aliteq