FADU FS-SSD Brings File-Aware Storage Closer to GPUs and DPUs

FADU’s File-Semantic SSD, or FS-SSD, is not being pitched as just another faster drive. Announced at the 2026 OCP Global Summit and reported on October 21, 2026, it takes a bigger swing at the storage stack: moving file-system awareness into the SSD so accelerators can work with files without leaning on the host CPU in the usual way. In FADU’s model, GPUs, DPUs, and NICs can open and read files without a host file system, block layer, or storage-server CPU sitting directly in the path. That makes FS-SSD interesting for AI, content delivery, JBoF storage, and other setups where the server CPU is increasingly becoming a control-plane bottleneck rather than the center of the data path. (tomshardware.com)

Traditional SSDs expose blocks, while the host operating system handles file lookup, permissions, metadata, caching, and the translation between files and storage. FS-SSD shifts part of that work into the drive. According to the details reported so far, FADU uses NFSv3 operations carried inside standard NVMe commands, with the drive handling path lookup, access control, and file-to-flash mapping. Under the hood, the file layer is derived from F2FS, while the file system and flash translation layer are merged more tightly than they would be in a normal host-managed setup. The idea is not to replace every data-center file system overnight, but to give accelerator-led systems a way to request files more directly and avoid dragging every operation through a conventional CPU-heavy storage stack. (tomshardware.com)

See Crucial T705 13 GB/s NVMe SSD price

The timing makes sense. Modern AI and analytics servers often have expensive accelerators waiting on data movement, metadata handling, or coordination outside the GPU. FADU says file-system and block-layer work can consume roughly 30% of CPU cycles in a GPUDirect Storage random-read workload, which is the kind of overhead FS-SSD is trying to reduce. The drive is also designed for scenarios where a GPU builds NFS-over-NVMe commands in its own memory, or where a SmartNIC or DPU places commands into the SSD submission queue. Potential use cases named by FADU include GPU file reads and checkpoints, DPU-served HTTP content delivery, and a JBoF file server where flash sits behind a PCIe switch and accelerator-friendly networking hardware. (tomshardware.com)

View NVIDIA DGX Spark Personal AI Desktop Supercomputer on partner website

The headline number is eye-catching, but it needs context. In FADU’s preliminary NFS-over-NVMe testing on a drive built around its FC6161 Gen6 controller, remote file throughput reportedly stayed around 460,000 ops/s while the server was reduced from 16 cores to one. The same report notes that Linux NFS-over-RDMA was much lower in that comparison, dropping from roughly 150,000 ops/s with 16 cores to around 40,000 ops/s with one core. However, those FS-SSD tests were run with the Gen6 controller in a PCIe 5.0 host, not a PCIe 6.0 host, and FADU’s Gen6-host results are still expected later. So the useful takeaway is not that the final product is fully proven, but that the architecture shows why moving file semantics closer to flash could matter. (tomshardware.com)

Buy SANDISK Optimus GX PRO 8100 PCIe 5 SSD here

As of Thursday, October 22, 2026, FS-SSD is still an active-development technology rather than a product broadly available to deploy. FADU expects Gen6-host results by Q1 2027 and partner samples in Q3 2027. There are also real system-design questions around locking, credentials, ACLs, caching, crash handling, multi-host coordination, and how GPU, DPU, and host writers avoid stepping on one another. FADU has described a model where some coordination remains with clients or metadata servers, while capability-based authentication and future Kerberos support are part of the security direction. That makes FS-SSD less of a simple plug-in upgrade and more of a glimpse at where high-performance storage may be headed: drives that are not just capacity endpoints, but active participants in AI and accelerator architecture. (tomshardware.com)

See ASUS GeForce RTX 5090 price