Nvidia open-sources cuFile API, accelerating GPU read/write capability for high-speed storage
SiliconANGLE Kyt Dotson ● Covered by 3 sources
Nvidia open-sourced cuFile, letting GPUs pull data from storage in milliseconds instead of waiting on the CPU. That matters because idle GPUs burn cash — this closes AI's biggest data bottleneck.
Based on reporting by SiliconANGLE, Kyt Dotson — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nvidia used the Future of Memory and Storage conference to announce it's open-sourcing cuFile, the API that sits at the heart of its GPUDirect Storage stack. CuFile itself isn't new — it hit general availability back in July 2021 alongside CUDA Toolkit 11.4 — but making the API open changes who gets to build on top of it. The pitch is simple: millisecond-level access from storage straight into GPU memory, no detours required.
The technical trick is direct memory access, or DMA, which lets data move from something like an NVMe drive right into GPU memory without routing through the CPU or system RAM first. That routing has always been the slow part. Cut it out and you avoid what Nvidia calls the GPU starvation loop — the situation where GPUs in a cluster sit there doing nothing because the data pipeline can't keep up with them. For retrieval-augmented generation and agentic AI workloads, which lean on huge datasets and expect fast responses, that gap between storage and compute is exactly where performance dies.
Alongside the open-sourcing, Nvidia rolled out Storage-Next, a broader industry push involving roughly 40 storage and flash memory vendors, DataDirect Networks, Kioxia and Micron among them. The idea is to get storage makers, controller vendors, cooling specialists and standards bodies pulling in the same direction on GPU-driven storage. DDN's chief technology officer, Sven Oehme, framed it as a shift in how AI success gets measured — not by how much infrastructure a company owns, but by how well it's actually used, with a tighter GPU-to-data connection as the mechanism.
Nvidia is building this around something called SCADA — scaled, accelerated data access — meant to give massively parallel GPUs high-bandwidth, low-latency pulls at scale. But speed like this comes with a catch. Letting AI systems reach storage media directly can punch through the usual safeguards against clobbering, where two processes overwrite each other's work, and can also sidestep encrypted or privileged access controls that were there for a reason. Nvidia's fix splits the job in two: raw, fast access lives outside the secure computing base for speed, while a separate privileged layer handles protected access under standard Linux rules to keep things from colliding or leaking.
My take — AI-written commentary, not fact-checked reporting
Open-sourcing the plumbing while keeping the real leverage — Storage-Next, SCADA, the vendor coalition — firmly Nvidia-shaped is a familiar move, and a smart one. It buys goodwill and adoption from the API release without giving up control of where the ecosystem actually goes. The security tradeoff here deserves more scrutiny than a press announcement will give it: any system that punches through clobber protection and privileged access controls for the sake of speed is one bad implementation away from a real headache, and burying that in a two-tier access model doesn't make the risk disappear, it just relocates it.
Read more about this at: SiliconANGLE