Nvidia open-sources cuFile API, accelerating GPU read/write capability for high-speed storage
SiliconANGLE Kyt Dotson ● Covered by 3 sources
Nvidia just made its cuFile storage API open-source, letting GPUs grab data straight from drives without waiting on the CPU. That means less GPU idle time and faster AI training on massive datasets.
Nvidia is giving away the keys to one of its quieter but crucial pieces of AI plumbing. The company announced today that it's open-sourcing the API for cuFile, the technology underpinning its GPUDirect Storage stack, which lets GPUs pull data directly from NVMe drives into their own memory. No CPU pit stop, no system memory detour. Just a straight line from storage to silicon, with access times measured in single-digit milliseconds.
The problem this solves has a name inside Nvidia's world: GPU starvation. Picture a rack full of $30,000-plus accelerators sitting there, burning power, waiting for data that hasn't arrived yet. That's the nightmare scenario for anyone running retrieval-augmented generation or agentic AI workloads, where models need to yank huge chunks of data on demand and fast. CuFile has been around since 2021, tucked into CUDA Toolkit 11.4, but keeping it proprietary meant only Nvidia controlled how storage vendors could plug into it. Open-sourcing the API flips that.
Alongside the code release, Nvidia unveiled Storage-Next, an alliance of roughly 40 companies spanning flash memory, controllers, cooling and orchestration — DataDirect Networks, Kioxia and Micron among them. The goal is blunt: standardize how storage hardware talks to GPUs so the whole industry isn't reinventing bottleneck fixes independently. DDN's CTO Sven Oehme put it plainly, arguing that AI success increasingly hinges on how efficiently companies use the infrastructure they already have, not how much more they buy.
The technical backbone here is something Nvidia calls SCADA — scaled, accelerated data access — and it's solving a problem that sounds almost mundane until you think about the scale involved. When multiple processes try writing to the same data simultaneously, you get clobbering, one overwriting the other's work. Direct storage access at GPU speed makes that risk worse, not better, since it can also sidestep encryption and privileged-access checks that normally guard against exactly this kind of collision. Nvidia's fix splits the job: raw, fast access for the application layer, kept outside the trusted security perimeter, while a separate privileged layer enforces Linux-style protections to keep things from stepping on each other.
What's notable is the timing. Mixture-of-experts models now dominate serious inference workloads, and they're voracious about pulling data mid-thought to fire off tool calls and reasoning chains. Nvidia is betting that the next competitive edge in AI infrastructure won't come from faster chips alone, but from making sure those chips are never sitting around waiting.
My take
Nvidia doesn't open-source things out of generosity; it does it when standardizing the plumbing benefits Nvidia's chips more than it benefits any single storage vendor. This is smart, not selfless — get 40 companies building tooling around your GPU access patterns, and you've locked in the ecosystem without needing exclusivity clauses. The security bolt-on for clobbering and encryption bypass is the part to watch, because bypassing safeguards for speed always creates a debt somebody eventually pays, usually a customer who assumed 'fast' also meant 'safe' by default.
Read more about this at: SiliconANGLE
