FADU’s FS-SSD design puts file handling inside the drive

FADU’s File-Semantic SSD (FS-SSD) is designed to handle file operations inside the drive, using NFSv3 operations carried in standard NVMe commands. That could let supported GPUs, DPUs and network cards request files directly, while a kernel-mounted client still uses host software. FADU’s early performance figures are preliminary company results, not results from a PCIe 6.0-host test.

What FS-SSD does inside the drive

In FADU’s design, the SSD handles file-path lookup and maps files to flash. Its file layer derives from F2FS, a Linux file system, and is co-designed with the flash translation layer (FTL), the part of an SSD that maps logical data to physical flash. FADU says this lets the drive coordinate file mapping and flash-management work internally.

The design carries 21 NFSv3 operations inside standard NVMe commands, rather than introducing a new file protocol. FADU describes support for PCIe and for NVMe over Fabrics (NVMe-oF) over RDMA or TCP. It also says the design allocates about 0.1% of NAND capacity to DRAM metadata, or roughly 1 GB per terabyte of NAND, with about 95% of that DRAM available for a frequently used node cache.

Direct access and kernel clients take different paths

The host-free path applies to direct access from supported accelerators or network devices, not every way of using FS-SSD. A GPU can submit NVMe commands from queues in its own memory; a DPU or SmartNIC can submit them through NVMe-oF offload. A kernel-mounted client still runs host software and incurs kernel overhead.

This difference matters when reading FADU’s performance figures. A result from a direct SPDK client, a software framework for storage applications, does not describe the same access path as a kernel client.

What FADU’s preliminary results show

In its October 12, 2026 white paper, FADU reported about 460,000 operations per second in a preliminary remote Filebench test, with throughput staying near that level as the server allocation fell from 16 cores to one. The test used an FC6161 Gen6 controller in a PCIe 5.0 host. FADU’s Linux NFS-over-RDMA comparison measured about 150,000 operations per second at 16 cores and about 40,000 at one core.

FADU also reported preliminary 4 KB random-I/O results using four threads, a queue depth of 128 and direct I/O. The table shows the reported read and write throughput for each setup. These results use the same workload size and units, but the access methods differ.

From FADU’s figures, the FS-SSD SPDK client delivered about 1.43× Ext4’s random-read throughput and 1.54× its random-write throughput. The kernel client’s results were lower than Ext4’s in both measures: 8,070 versus 9,120 MB/s on reads, and 4,320 versus 4,560 MB/s on writes. The raw-block SPDK random-write figure was 13,240 MB/s; FADU’s cited NAND program bandwidth was about 12 GB/s, and that result may not represent steady-state performance.

FADU’s reported 63% reduction in write amplification versus host F2FS came from a mixed sequential-and-random workload run on the NVMeVirt software SSD emulator.

Write coordination and security

The described FS-SSD implementation has no cross-host locking. If separate hosts write to the same location, writes are applied whole and in order, with the later conflicting write taking precedence. GPU and SPDK clients must coordinate at the application level; a pNFS metadata server can coordinate access at scale.

FADU describes capability-based authentication and says Kerberos support is planned. The reported feature set does not include ACLs, or access-control lists, and other file-system features. The FC6161 controller’s stated encryption support is TCG Opal self-encryption.

FADU’s targets for 2027

FADU’s roadmap, reported in October 2026, targeted Gen6-host results by Q1 2027, partner samples in Q3 2027 and a hostless JBoF reference design by the end of 2027. Those are planned milestones.

For the proposed JBoF design, FADU’s model estimates about 42% lower power and 57% lower cost per 16-drive server, roughly 1.7× the rack density and 60% lower total cost of ownership over five years. These are projections from FADU’s model.