# FADU Proposes File-Semantic SSDs to Bypass Host CPU Bottlenecks FADU aims to integrate file systems directly into SSD firmware to accelerate AI workloads. - **Source:** [Latest from TechRadar](https://www.techradar.com/feeds.xml) - **Published:** 2026-10-29 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/a65daf8c-fa48-4733-9620-199202a8c297 ## Redefining Storage Path Architecture FADU has introduced the File-Semantic SSD, which shifts file lookup, permission checks, and file-to-flash mapping directly into the controller firmware. By moving these tasks off the host CPU, the design aims to eliminate the traditional block layer and file system overhead that currently creates significant bottlenecks in high-performance AI storage environments. ## Leveraging NFSv3 for Compatibility Rather than creating a proprietary protocol, FADU utilizes NFSv3 to handle file requests within standard NVMe commands. This stateless approach allows the drive to process requests independently without needing to track open file states. The implementation supports both local server access and networked NVMe-oF configurations, ensuring broad compatibility with existing data center infrastructures. ## Optimizing Flash Translation Layers The FS-SSD merges file system mapping and flash translation into a single layer based on F2FS. This integration allows the drive to map files directly to physical flash locations. By optimizing this map, FADU claims 95% of the drive's DRAM can be repurposed to cache frequently accessed file details rather than block maps. ## Reducing Write Amplification Impact By consolidating garbage collection into a single process, FADU estimates a 63% reduction in write amplification compared to standard host-based F2FS implementations. This reduction significantly lowers wear on flash memory cells. However, these figures are currently derived from simulated tests using NVMeVirt software rather than measurements from production-ready physical hardware drives. ## Performance Metrics and Limitations Preliminary tests on a PCIe 5.0 host showed the FS-SSD maintaining 460,000 ops/s even when server core counts were reduced. While impressive, these results are constrained by the PCIe 5.0 link limit of 14.1GB/s. The company has yet to demonstrate performance on Gen6 hardware, which is already reaching speeds of 28GB/s in mass production. ## Competitive Landscape and Precedents FADU is not the first to address CPU-intensive storage paths; Nvidia’s GPUDirect Storage previously removed the CPU from the data path. While Nvidia shifts control to the GPU, FADU’s approach focuses on the drive itself. Previous research like the 2018 DevFS paper also explored embedding file systems directly into storage devices. ## Future Roadmap and Projections FADU plans to release partner samples by Q3 2027, with a reference design using Xsight Labs’ 800G E1 DPU expected by the end of that year. The company projects a 42% reduction in power consumption and 57% lower costs for 16-drive systems, though these remain theoretical models rather than verified operational data. ## Strategic Significance for AI The success of this architecture hinges on the Q1 2027 PCIe 6.0 host results, which will prove if the drive can handle CPU tasks at full speed. As AI storage demands grow, FADU’s ability to standardize this command set through bodies like NVM Express will determine its long-term viability as a storage solution. ## Sources covering this story - https://www.techradar.com/pro/fadu-wants-to-take-the-cpu-out-of-ai-storage-by-putting-the-file-system-inside-the-ssd-but-the-proof-wont-arrive-until-2027 — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._