Korean Controller Maker FADU Unveils 'File-Semantic' SSD to Offload File System Logic from Host CPU
FADU showcased FS-SSD at the OCP Summit, embedding NFS requests in firmware to bypass host CPU bottlenecks, though performance data remains largely simulated or early-stage, with samples due Q3 2027.
South Korean storage controller designer FADU has unveiled its File-Semantic SSD (FS-SSD) at the OCP Global Summit. The drive aims to address a bottleneck in AI data centers: in traditional architectures, the host CPU spends significant time handling file lookups, permission checks, and address mapping. FADU's solution embeds these operations directly into the firmware of its sixth-generation controller, allowing GPUs, DPUs, and network cards to access files without passing through the host file system or block layer.
Technically, FS-SSD borrows the mature Network File System protocol NFSv3, encapsulating its requests within standard NVMe commands. FADU states this design leverages the stateless nature of NFSv3 to simplify drive memory management. Additionally, the drive merges two lookup processes—normally handled separately by the file system and the flash translation layer—using a layout based on Linux's F2FS file system, aiming to reduce DRAM usage and optimize garbage collection efficiency.
FADU published preliminary performance figures, but they require careful interpretation. The company claims that in GPUDirect Storage random-read workloads, the host file system and block layer consume about 30% of CPU cycles. In remote file throughput tests, FS-SSD maintained approximately 460,000 ops/s as server cores were reduced from 16 to one, while a standard Linux NFS server dropped from roughly 150,000 ops/s to 40,000. However, these tests were conducted on a PCIe 5.0 host, and key metrics like a 63% reduction in write amplification come from simulations using NVMeVirt software developed by Seoul National University, not real hardware.
Currently, FS-SSD is not an industry standard; FADU has not submitted a file-semantic command set to NVM Express, so clients still require proprietary software. Partner samples are not expected until Q3 2027, when real-world testing in PCIe 6.0 environments will be crucial to verify if this architecture can truly alleviate AI storage bottlenecks.