FADU Puts the File System Inside the SSD to Bypass the CPU

Embedding NFS directly into NVMe commands

FADU encapsulates stateless NFSv3 calls inside regular NVMe commands to eliminate the host operating system from the storage pipeline.

Modern enterprise storage drives are block devices. They do not know what a file is; they only understand raw sector addresses, leaving the host operating system to run the file system, resolve directories, and track permissions. In dense AI server clusters where thousands of GPUs hammer storage arrays simultaneously, managing those lookups consumes substantial processing time.

FADU claims that in typical random-read workloads using Nvidia GPUDirect Storage, the host file system and block layer chew through roughly 30% of the storage server's CPU cycles. While GPUDirect handles data movement directly between drives and GPU memory, the host CPU still has to orchestrate every lookup.

To bypass that bottleneck entirely, FADU's File-Semantic SSD (FS-SSD) embeds an internal file management layer based on Linux's Flash-Friendly File System (F2FS) directly inside its sixth-generation controller. The drive speaks Network File System version 3 (NFSv3), packaging standard file requests inside regular NVMe commands. Because NFSv3 is lightweight and stateless, the SSD processes directory lookups and permission checks on its own silicon without tracking connection states.

FADU File-Semantic SSD specifications and test results

Consolidating maps and slashing write amplification

In a conventional server, two separate systems translate storage locations. The host file system converts a file name into logical block numbers, and the SSD's internal Flash Translation Layer (FTL) maps those blocks to physical NAND cells. FS-SSD merges both steps into a single map organized around the physical geometry of the flash.

That direct mapping frees up internal resources. FADU notes that while an enterprise drive typically dedicates almost all of its DRAM cache (about 1GB per terabyte of flash) strictly to logical-to-physical block tables, its unified file layer leaves around 95% of that DRAM free to cache frequently accessed file metadata.

Coordinating cleanup routines also saves NAND health. Because conventional operating systems and SSD controllers run garbage collection independently, their background operations often clash and rewrite the same sectors. FADU claims its consolidated design cuts write amplification by 63% based on simulated tests using Seoul National University's NVMeVirt environment, significantly extending drive lifespan under heavy write cycles.

Early benchmarks show promise, but hardware is years away

In preliminary tests presented by FADU, remote file throughput held steady at roughly 460,000 operations per second even as the storage server was throttled down from 16 cores to just one. By comparison, a conventional Linux NFS configuration under identical conditions dropped from around 150,000 ops/s to 40,000.

Those numbers come with major caveats. The test rig ran desktop-class Intel Core i5-13600K processors rather than enterprise server silicon, and the controller was limited by a PCIe 5.0 interface capped at 14.1 GB/s. With native PCIe 6.0 enterprise drives already entering mass production, raw interface throughput will matter just as much as command latency when training modern clusters, as seen in modular deployments like the Nvidia DSX POD architecture.

The largest catch is the calendar. While FADU plans to demonstrate PCIe 6.0 benchmark data in early 2027, actual partner samples of the FS-SSD will not ship until the third quarter of 2027. Moving the file system inside the drive is a technically elegant solution to CPU bottlenecks, but data center architects will be waiting almost a full year to test physical hardware.

Availability and pricing

FADU has not announced pricing or commercial product names for drives powered by its Gen6 FS-SSD controller. The architecture was shown as a technology demonstration targeting hyperscalers and data center operators rather than retail buyers.

Partner sampling across North America and Europe is slated for the third quarter of 2027, with volume production expected later depending on qualification cycles.