# FADU’s File-Semantic SSD Moves the Filesystem Into the Drive, So GPUs Can Skip the CPU > Source: > Published: 2026-10-21 18:11:12+00:00 **FADU has built a file-semantic SSD that runs the filesystem on the drive itself, so GPUs, DPUs and network cards can open and read files without a host CPU in the path. The early numbers are striking, and so are the caveats.** Every SSD you have ever bought speaks in blocks. The drive knows nothing about files, folders or permissions; it hands out numbered chunks of flash, and the operating system’s filesystem turns those into something an application can use. That arrangement has worked for decades. In an AI server where the GPUs are hungry and the CPU is mostly a traffic cop, though, it is starting to look like overhead. FADU, the Seoul-based fabless company behind the controllers in drives such as the SN861 that Western Digital launched and SanDisk now sells, used the OCP Global Summit in San Jose to show its answer: a file-semantic SSD, which it calls FS-SSD. The drive handles path lookup, access control and the mapping from file to flash address on its own, with firmware running on FADU’s sixth-generation FC6161 controller. ## NFS, tucked inside NVMe The trick is in the protocol. Rather than invent a new transport, FADU wraps NFSv3 operations inside ordinary NVMe commands. [According to Tom’s Hardware](https://www.tomshardware.com/pc-components/ssds/fadu-puts-the-filesystem-inside-the-ssds-flash-to-bypass-host-cpus-file-semantic-pcie-6-0-ssd-delivers-460-000-ops-s-to-gpus-and-dpus-samples-due-by-q3-2027), NFSv3 was picked because it is small, stateless and indifferent to the operating system, which keeps the firmware lean. It works over PCIe today and over NVMe-oF on RDMA or TCP without changes. Under the hood, the file layer is derived from F2FS and merged with the flash translation layer, so there is one garbage collector instead of two fighting each other. A GPU can build file requests in its own memory, and a SmartNIC or DPU can drop them straight into the drive’s queue, with data moving peer-to-peer over PCIe. FADU’s case for doing this is a measurement of its own: in a GPUDirect Storage random-read workload, the filesystem and block layers ate roughly 30% of CPU cycles. ## The numbers, and what is behind them The headline figure comes from a Filebench NFS server test. FS-SSD held about 460,000 operations per second whether the server had 16 cores or one. Linux NFS over RDMA managed around 150,000 ops/s with 16 cores and about 40,000 with a single core. In fio, sequential writes through the kernel client reached 11,940 MB/s, well ahead of Ext4 at 7,050 MB/s on the same hardware, while random reads landed slightly behind it. Read the fine print before getting excited. The test rig was a Core i5-13600K desktop with a 100 Gbps ConnectX-6 card, and the Gen6 drive sat on a PCIe 5.0 host, which is why sequential reads stall near 14.1 GB/s. Against raw block access through SPDK, the in-drive file layer costs about 7% on 4K random reads and close to half on 4K random writes. The 63% write-amplification reduction FADU quotes comes from a software emulator built at Seoul National University, not from silicon. There is also no cross-host locking yet, last writer wins, and features such as snapshots, quotas and ACLs are not described. ## Where it goes from here FADU’s longer play is a hostless storage box: a shelf of FS-SSDs behind a PCIe switch, with a DPU handling the network, and no server CPU at all. Its own model puts that at roughly 42% lower power and 57% lower cost per 16-drive server, numbers that should be treated as a pitch until someone builds one. The [Korea Times](https://www.koreatimes.co.kr/business/companies/20261015/fadu-unveils-worlds-1st-file-semantic-ssd-to-power-ai-workloads) reports the demo used a 16 TB drive and that a complete Gen6 SSD stays under 28 W in sequential writes. The timeline is not short. Gen6-host results are due by Q1 2027, partner samples in Q3 2027, and a JBoF reference design on Xsight Labs E1 DPUs by the end of that year. Nothing has been submitted to NVM Express or OCP as a standard, and that is the part that will decide whether this stays a clever FADU feature or becomes something every controller vendor has to answer. Desktop users will not see it any time soon. The datacenter crowd, short on CPU cycles and power budget, is the audience, and it will want those Gen6 numbers first.