FADU, a fabless SSD controller designer based in Seoul, South Korea, has announced a File-Semantic SSD (FS-SSD) at the 2026 OCP Global Summit. The device moves the file system into the drive itself, so that GPUs, DPUs and network cards can open and read files with, in the words of the company's white paper, "no host filesystem, no block layer, and no storage-server CPU in the path." FADU told Tom's Hardware Premium it expects partner samples by the third quarter of 2027.

The company says the approach targets the control path, since the data path moved to accelerators years ago. According to FADU's measurements, the file system and block layers can consume about 30% of CPU cycles in a GPUDirect Storage random-read workload. Existing bypass options, it notes, give up files altogether, while its design aims to keep both files and bypass.

In a preliminary test of NFS over NVMe on a drive built around its FC6161 Gen6 controller, installed in a PCIe 5.0 host, remote file throughput stayed at roughly 460,000 operations per second as the server was scaled from what FADU labels 16 cores down to one. Linux NFS-over-RDMA reached about a third of that figure with 16 cores and only 40,000 operations per second with a single core. FADU describes the technology as under active development and expects results on a Gen6 host by the first quarter of 2027.

NFSv3 operations travel inside standard NVMe commands, and the drive handles path lookup, access control and the mapping of files to flash. FADU chose NFSv3 because it is small, with 21 operations, agnostic to operating system and file system, and stateless. Internally, the file system and flash translation layer are merged, using a file layer derived from F2FS whose node table becomes a "File Map" pointing to physical flash addresses. The drive uses about 1 GB of DRAM per terabyte of NAND, as a standard SSD does, but roughly 95% of it can serve as a node cache. A single garbage collector replaces the two used with host F2FS, and FADU reports 63% lower write amplification on a mixed workload. That figure comes from a research prototype running on NVMeVirt, an open-source SSD emulator from Seoul National University, so it is not a hardware result.

In operation, a GPU can build NFS-over-NVMe commands in queues in its own memory, or a remote SmartNIC or DPU can place them in the drive's submission queue. Data moves from the SSD to the network card by PCIe peer-to-peer DMA and leaves as RDMA writes. FADU's named use cases are GPU file reads and checkpoints, DPU-served HTTP content delivery and a just-a-bunch-of-flash (JBoF) file server.

The test platform used Intel Core i5-13600K client and server machines with 128GB of DDR5 and a Mellanox ConnectX-6 100 Gbps NIC. Because the host is PCIe 5.0, sequential reads top out near 14.1 GB/s, roughly the link limit. Against SPDK on the raw drive, FS-SSD's SPDK client keeps about 93% of 4K random-read performance and 53% of 4K random-write performance. The kernel client trails Ext4 by 11.5% on random read and 5.3% on random write, by Tom's Hardware's calculation.

Several questions remain. FADU said there is no cross-host locking and that writes follow a last-writer-wins rule, with coordination left to clients or a pNFS metadata server. To stop clients from forging user and group IDs, it uses capability-based authentication, with each command carrying a capability from an outside server; Kerberos support is planned. ACLs and several other features are absent from the current release, encryption relies on the controller's TCG Opal support, and capacitors protect data on power loss.

Looking further ahead, FADU envisions a JBoF design with no front-end server, built from FS-SSDs, a PCIe switch and a DPU or NIC. Its model projects about 42% lower power and about 57% lower cost per 16-drive server. Real numbers, samples and a JBoF reference design are expected next year, and Tom's Hardware says a Gen6-host run is now needed to support the company's claims.