NVIDIA just released SCADA, letting GPU bypass CPU and connect directly to hard drives, speeding up 5x
NVIDIA Just Released SCADA, Letting GPUs Bypass the CPU to Connect Directly to Storage—Speeds Up to 5x Faster
At the FMS 2026 Flash Memory Summit on August 5, NVIDIA unveiled a new architecture—SCADA—and fully open-sourced it, completely opening up the data path for GPUs to connect directly to SSDs, with no more CPU middleman in the way.
In the old days of data transfer, the CPU was an unavoidable relay station
When a GPU needed to read data, it had to go through
💡 What You Will Learn
NVIDIA Just Released SCADA, Letting GPUs Bypass the CPU to Connect Directly to Storage—Speeds Up to 5x Faster At the FMS 2026 Flash Memory Summit on August 5, NVIDIA unveiled a new architecture—SCA
📜 Table of Contents
NVIDIA Just Released SCADA, Letting GPUs Bypass CPUs to Connect Directly to SSDs — 5x Faster
At the FMS 2026 Flash Memory Summit on August 5th, NVIDIA unveiled a new architecture — SCADA — and fully open-sourced it, completely clearing the data path for GPUs to connect directly to SSDs, with no more CPU middleman.
In the Old Data Pipeline, the CPU Was an Unavoidable Relay Station
When a GPU needs to read data, the traditional flow goes like this: SSD → CPU memory → GPU memory. The CPU not only had to play the role of mover, but also handle all the scheduling and command coordination.
NVIDIA introduced GPUDirect Storage (GDS) back in 2019, which allowed data to bypass CPU memory and transfer directly to the GPU. But GDS only solved the data path problem — the command and control still had to go through the CPU. It worked fine for reading large files, but when faced with AI inference workloads generating massive numbers of 512-byte small requests, CPU scheduling became the bottleneck.
To use an analogy: GDS is like building a dedicated express highway for cargo, but the traffic lights are still managed by the CPU. SCADA simply hands the traffic lights over to the GPU itself.
SCADA Hands Storage Scheduling Control to the GPU
SCADA is a new storage architecture defined by NVIDIA, and the core change boils down to one sentence: The GPU directly initiates storage I/O requests and manages the NVMe drive queue depth itself. The CPU is completely out of the picture.
The benchmark data is striking: in data analytics workloads, SCADA mode is 5.3x faster than the traditional CPU-triggered mode, and hardware costs can be reduced by up to 21.7x.
Micron showcased a reference design on-site, using 3 H100 GPUs to drive 44 PCIe Gen6 SSDs, achieving 230 million random read IOPS and 118GB/s of storage throughput on a single server. The CPU responsible for scheduling remained nearly idle throughout the entire process.
Meanwhile, NVIDIA fully open-sourced its GPU storage acceleration API, cuFile. Intel has also joined as an open-source maintainer, meaning cuFile and SCADA have transitioned from NVIDIA's proprietary technology into industry-wide standards.
A 40-Vendor Alliance to Rebuild Storage for the GPU Era
KV cache in AI inference constantly generates small data reads and writes. Traditional SSDs reading with 4KB sectors suffer from an 8x read amplification problem. To address this, NVIDIA established the Storage-Next industry alliance, focusing on optimizing read/write performance for 512-byte sectors.
The alliance now has over 40 members, including DDN, Kioxia, Micron, Samsung, as well as storage controller companies, thermal solution vendors, and standards organizations. Together, they aim to define a new set of standards: how SSDs should adapt when GPUs — not CPUs — become the primary initiators of storage access.
On the hardware front, NVIDIA has built the STX architecture system based on the Vera BlueField-4 storage processor, with commercial systems planned for the second half of 2026.
Does This Technology Matter for Gamers?
Many people will ask: does GPU-direct-to-SSD connectivity help gaming? The answer is: not directly right now, but the underlying concept is the same.
SCADA and GDS currently only support Linux, not Windows. And NVIDIA's technology is aimed exclusively at data center-grade hardware (Tesla/Quadro cards) — GeForce gaming cards can't use it.
That said, Microsoft has already done something similar — DirectStorage was designed specifically for gaming, allowing NVMe SSDs to feed textures and model data directly to the GPU. And SCADA's strength lies in 512-byte small packets (AI inference scenarios), whereas games load large chunks of data — it's not the same pain point.
But if game engines start integrating real-time AI inference in the future (think AI NPCs, real-time scene generation), SCADA's low-latency advantage would come into play. GPU-direct storage will eventually make its way from the data center into your gaming PC.
A true AI infrastructure upgrade isn't just about growing compute power — storage throughput and data movement need to keep pace too. SCADA fills in a critical piece of the AI storage puzzle.
Do you think GPU-direct storage technology will change the hardware architecture of personal computers in the future? Feel free to share your thoughts in the comments.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
