FADU FS-SSD Moves File System Into Flash: 30% CPU [2026]

FADU introduced the File-Semantic SSD (FS-SSD) at the OCP Global Summit, a drive built on FADU's sixth-generation controller whose firmware handles file lookup, access checks and file-to-address translation, carrying NFSv3 requests inside standard NVMe commands over an F2FS-based internal file layer. FADU's preliminary test reported 460,000 operations per second as the server shrank from 16 CPU cores to one, and the company claims the host file system and block layer consumed about 30% of CPU cycles during random reads through GPUDirect Storage; no independent validation has been reported, and FS-SSD results in a PCIe 6.0 host are not expected until Q1 2027.

FADU wants to pull the host CPU out of the path between an AI server’s GPUs and its flash. At the OCP Global Summit, the company introduced the File-Semantic SSD FS-SSD , a drive that runs a file system inside its own controller firmware. TechRadar covered the pitch, and the headline-grabbing proof will not arrive until 2027. The design sits on FADU’s sixth-generation controller. Firmware on that chip handles file lookup, access checks, and file-to-address translation, the jobs a host operating system normally performs. FADU’s whitepaper says GPUs, DPUs, and network cards can reach files with “no host filesystem, no block layer, and no storage-server CPU in the path.” That sentence describes a design goal and a claimed access path, not a shipping product. The one performance figure so far comes from a preliminary test, and no independent validation has been reported. Buyers weighing storage for AI clusters, a spending line we examined in our look at Big Tech’s AI backlog against its capex https://shattered.io/big-tech-earnings-1-7t-backlog-vs-600b-capex-ai-test-2026/ , should treat FS-SSD as something to track through 2027 and not something to order today. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F What FADU Announced at the OCP Global Summit FADU used the OCP Global Summit to put a name on an idea that storage engineers have circled for years: let the drive understand files, not just blocks. The product is the File-Semantic SSD, shortened to FS-SSD. It is built around FADU’s sixth-generation controller, and the firmware carries the file-system logic that would otherwise live in the host kernel. Reporting on the announcement, including TechRadar’s coverage, points to a handful of concrete details. Here is the short version of what was stated publicly. - Product: File-Semantic SSD FS-SSD , shown at the OCP Global Summit. - Controller: FADU’s sixth-generation design. - Firmware role: file lookup, access checks, and file-to-address translation inside the drive. - Protocol: NFSv3 requests carried inside standard NVMe commands. - Internal file layer: based on F2FS, a Linux file system built for flash memory. - CPU claim: the host file system and block layer ate about 30% of CPU cycles during random reads through GPUDirect Storage. - Preliminary result: 460,000 operations per second as the server shrank from 16 CPU cores to one. - Next milestone: FS-SSD results in a PCIe 6.0 host by Q1 2027. Everything on that list comes from FADU’s own materials as relayed by the press. None of it has been checked by a third-party lab. That distinction matters for the rest of this analysis, because a vendor whitepaper and a validated benchmark are different kinds of evidence. How the File-Semantic SSD Works A normal NVMe SSD speaks in blocks. The host decides which logical block addresses hold which bytes of which file, and the drive obeys. That bookkeeping happens in the host file system and the block layer, both of which burn CPU time on every request. FS-SSD moves the bookkeeping across the bus and into the drive. NFSv3 requests inside NVMe commands FADU’s approach carries NFSv3 requests inside standard NVMe commands. The client asks for a file by name and handle, the way an NFS client would, and the drive resolves it. NFSv3 is an old and simple network file protocol, so it is easy to implement in firmware and familiar to a huge range of clients. That is our reading of why FADU picked it, not a reason the company gave. Using standard NVMe as the carrier also means the drive can plug into existing PCIe infrastructure and existing driver stacks. A GPU, DPU, or network card that can issue NVMe commands can, in theory, ask for files directly. The reports do not say which vendors’ devices have been tested in that role. An F2FS-based file layer inside the controller Inside the drive, the file layer builds on F2FS, a Linux file system designed for flash. F2FS is log-structured, which suits NAND because flash prefers appending over overwriting in place. Starting from a proven open design is a sensible shortcut for a hardware company. It also ties FS-SSD’s behavior to a file system whose strengths and quirks the Linux community already knows. One consequence is worth flagging early. A drive already runs a flash translation layer that remaps addresses and cleans up stale data. Stack a log-structured file system on top and two layers of garbage collection can interact. FADU has promised hardware write-amplification figures in its Q1 2027 report, which suggests the company knows this is the question skeptics will ask first. The 30% CPU Cost FADU Says It Measured FADU’s motivating number is a 30% tax. The company said the host file system and block layer consumed about 30% of CPU cycles during random reads through GPUDirect Storage. That is the headline justification for the whole project, so it deserves a close look. GPUDirect Storage is Nvidia’s technology for moving data between storage and GPU memory without bouncing it through system RAM. It removes the data copy from the CPU’s plate. It does not remove the CPU’s metadata work. The host still has to open the file, find its extents, check permissions, and build the block requests. FADU’s argument is that this leftover work is big enough to matter at AI scale. Two cautions apply. First, the figure covers random reads, which are metadata-heavy and a worst case for the host stack. Large sequential reads, such as streaming a checkpoint, would likely show a smaller share. Second, the number comes from FADU and the public reporting does not describe the test machine, kernel version, or file system used as the baseline. A different baseline could move the percentage in either direction. Even so, the direction of the claim is plausible. Anyone who has profiled a busy NVMe server has seen kernel storage code show up high in the CPU profile. The open question is how much of that 30% an in-drive file system truly removes, and how much simply moves to a different place on the same board. The 460,000 Operations Per Second Test, Read Carefully The only performance result on the table is a preliminary test that reached 460,000 operations per second while the server configuration was reduced from 16 CPU cores to one. Cutting fifteen cores out of the picture is a striking setup, and it is clearly meant to make the point that the host stops being the bottleneck. The framing implies that throughput held up as cores disappeared. The reporting we reviewed does not give a step-by-step curve, so we cannot say whether performance stayed flat, dipped, or even rose along the way. We also cannot say what counts as an operation here. That detail changes everything. Questions a reviewer would put to FADU: - Were the operations reads, writes, or a mix, and at what request size? - How many drives and how many GPUs or DPUs drove the load? - What was the queue depth, and did the single core still run the application? - What did the 16-core baseline deliver on the same hardware with a conventional file system? - Was the test run on a PCIe 6.0 host, or on an earlier generation? Until those answers appear, 460,000 operations per second is best read as a proof-of-concept data point. It shows the architecture works on a bench. It does not yet show the architecture wins in a rack. The Timeline: Q1 2027, Q3 2027, End of 2027 FADU laid out a staged path to evidence. The dates are targets stated by the company, so any of them can slip. The table below lays them in order and notes what each step would settle. | Milestone | FADU target | What it would settle | Status | |---|---|---|---| | Public debut at OCP Global Summit | Done | Concept, controller, protocol design | Announced | | FS-SSD testing in a PCIe 6.0 host | By Q1 2027 | Gen6 host behavior, hardware write-amplification figures | Planned | | Partner samples | Q3 2027 | Whether outside customers can reproduce the claims | Planned | | Reference design on an 800G E1 DPU | End of 2027 | Full path from network to flash without a host file system | Targeted | | OEM/ODM engagement | No date given | Route to server vendors and rack builders | Early-stage talks | The sequence is sensible. Gen6 host results come first because that is where the drive’s bandwidth gets stressed. Partner samples follow, which is the moment outside engineers can poke at the claims. The DPU reference design, reported for the end of 2027, closes the loop by showing the whole network-to-flash chain. It also means a buyer cannot test FS-SSD on their own workload before Q3 2027 at the earliest, and only if they become a sample partner. For 2026 and most of 2027, this is a roadmap story. What Is Confirmed and What Is Not Headlines about a CPU-free AI storage stack overshoot what the evidence supports. The table separates what the reporting backs from what it leaves open. | Claim | Standing | Basis | |---|---|---| | FS-SSD uses a sixth-generation FADU controller | Reported | FADU announcement | | NFSv3 requests ride inside NVMe commands | Reported | FADU whitepaper coverage | | File layer is based on F2FS | Reported | FADU materials | | Host file system and block layer use about 30% of CPU cycles in GPUDirect Storage random reads | Vendor-stated | FADU, not independently tested | | 460,000 operations per second at one CPU core | Preliminary, vendor-stated | FADU test, details not public | | FS-SSD is commercially released or broadly deployed | Unconfirmed | Samples are not due until Q3 2027 | | Independent validation of the quoted performance | Unconfirmed | None reported | | A price for FS-SSD | Unconfirmed | No price identified | | The CPU is eliminated from AI storage in every workload | Unconfirmed | Material describes a goal and a claimed path | The last row is the one to remember. FADU’s own language is about a data path where the host file system, block layer, and storage-server CPU are absent. Servers still need processors for provisioning, scheduling, monitoring, and running the training job itself. Removing the CPU from the storage data path is a narrower claim than removing the CPU from the AI server. Why AI Storage Has a CPU Problem Modern AI clusters pour money into accelerators, then surround them with ordinary servers whose CPUs shepherd data. Training jobs read huge datasets, write checkpoints, and pull model weights during inference start-up. Each of those flows crosses a storage stack that was designed when disks were slow and CPUs were the fast part. The mismatch now runs the other way. Flash is fast, GPUs are hungry, and the CPU in the middle is the part that has to keep up. Operators respond by adding cores, which costs power, rack space, and money. The shift to rack-scale GPU platforms, which we traced in our piece on Vera Rubin deployments across five clouds https://shattered.io/vera-rubin-five-clouds-aws-2-million-gpu-mix-2026/ , makes every wasted CPU cycle in the data path more expensive in relative terms. GPUDirect Storage was the first major fix. It cut the extra data copy. FADU is going after the second layer of overhead, the file-system and block-layer work that GPUDirect Storage leaves on the host. If the 30% figure holds up, that second layer is large enough to be worth attacking with new silicon. There is a cost angle too. Every core an operator does not buy for storage duty is a core available for other work, or a socket it can skip. The reported drop from 16 cores to one hints at the size of the prize, though real savings depend on what else those servers must do. Competitive Comparison: Five Ways to Feed GPUs FS-SSD enters a field where the file-system work already lives somewhere. The question is where, and who pays for the compute. The comparison below is qualitative, because FADU has not published numbers that line up against alternatives. | Approach | Where file logic runs | Host CPU role | Maturity | |---|---|---|---| | Conventional NVMe with host file system | Host kernel | Full: lookup, permissions, block mapping | Widely deployed | | GPUDirect Storage over a host file system | Host kernel | Metadata and block layer only FADU puts this at about 30% of cycles in random reads | Established | | Remote NFS or file appliance | Storage-server CPU | Moved to another server | Established | | DPU-offloaded storage | DPU processors | Reduced, DPU does protocol work | Growing | | FS-SSD | SSD controller firmware | Out of the data path, by FADU’s design goal | Pre-sample, tests due Q1 2027 | The DPU row is the most interesting pairing. FADU’s reported reference design uses an 800G E1 DPU, so the company sees DPUs as partners, not rivals. A DPU can terminate the network and hand file requests straight to the drive. In that picture the DPU and the SSD split the work that a server CPU used to do. On the accelerator side, the contest between rack platforms keeps raising the bar for data feeding. Our breakdown of AMD and Nvidia AI accelerators and their memory https://shattered.io/amd-vs-nvidia-ai-accelerators-mi455x-vs-vera-rubin-432gb-vs-288gb-2026/ shows how much on-package capacity vendors now pack in. Bigger local memory reduces how often a GPU must fetch from storage, but it does not erase the need to load datasets and checkpoints quickly. Storage still sets the pace at the edges of every job. Historical Context: Storage Keeps Pulling Compute Closer to Flash FS-SSD is the latest entry in a long line of attempts to make drives smarter. Computational storage proposals have tried to run filters and compression on the drive. Key-value SSDs tried to replace block addressing with object-style access. Zoned storage tried to push placement decisions up to the host. Each idea had real merit, and each struggled with the same obstacle: software ecosystems move slowly. What changed is the buyer’s pain. Hard drives and early SSDs were cheap enough that wasting host CPU did not hurt. Today the host CPU sits beside accelerators that cost far more, and idle accelerators are the new waste. Our comparison of HDD and SSD cost per terabyte and speed https://shattered.io/hdd-vs-ssd-12-25-per-tb-vs-7250-mbs-2026/ shows how far flash performance has outrun the software stack that serves it. FADU’s twist is the choice of interface. Earlier smart-drive efforts often invented new commands that needed new drivers. Tunneling NFSv3 through standard NVMe is a bet that compatibility wins adoption. If it works, operators will not need to rewrite their applications. They mount something that looks like a file system and let the drive do the rest. The same instinct shows up elsewhere in hardware. Vendors keep looking for ways to move work out of general-purpose cores. We saw it in the memory choices behind Intel’s LPDDR5X bet against HBM https://shattered.io/intel-d-series-lpddr5x-vs-hbm-480gb-memory-bet-2026/ , a different layer of the stack with the same logic: match the hardware to the data movement, not the other way around. Market Impact for Hyperscalers, OEMs, and Controller Vendors No price has been identified for FS-SSD, and partner samples are not expected until Q3 2027, so the market effect today is strategic and not financial. Still, the announcement sends signals to several groups. Hyperscalers and large AI operators are the obvious target. They run the clusters where a 30% CPU tax would add up to real money, and they hold the engineering depth to evaluate a new storage architecture. OCP is also their home turf, which explains the choice of venue. FADU said OEM and ODM talks are at an early stage, so server makers have not committed to anything yet. Controller rivals now face a choice. Either they ignore the file-semantic idea and compete on raw bandwidth, or they build their own offload story. Expect marketing language about data-path offload to spread across the SSD industry if FADU’s Q1 2027 numbers look good. DPU vendors get a potential tailwind. The reported 800G E1 DPU reference design links FS-SSD to network-side hardware, which would make DPUs more central to storage nodes. Software and file-system vendors face the opposite risk. If the file system migrates into drive firmware, the value of host-side storage software could shrink in the workloads FS-SSD wins. Supply conditions add a wrinkle. Memory is tight, as our coverage of IDC’s DRAM supply outlook https://shattered.io/ram-shortage-idc-16-percent-dram-supply-growth-2026/ explains, and a drive that tracks file metadata needs working memory to do it. FADU has not said how much DRAM or what capacities FS-SSD will use. That is analysis on our part, not a reported detail, but it is a cost variable buyers should ask about. Risks and Open Questions Putting a file system in firmware trades flexibility for speed. That trade creates a list of questions the company will need to answer before serious buyers commit. - Write amplification. A log-structured file layer on top of a flash translation layer can multiply internal writes. FADU’s promise of hardware write-amplification figures by Q1 2027 addresses this directly. - Security and tenancy. The drive performs access checks. Operators will want to know how those checks map to identities, how they handle multiple tenants, and how the firmware gets audited. NFSv3 itself offers limited built-in security, so the surrounding design matters. - Failure and recovery. File systems need crash consistency and repair tools. When the file system lives in firmware, a bug can mean a bricked namespace instead of a kernel patch. - Firmware updates. Fleet operators patch kernels often. They will ask how fast and how safely they can patch a file system that now ships inside the SSD. - Feature coverage. NFSv3 is simple. Snapshots, quotas, and rich permissions may not map cleanly, and the reports do not say what FS-SSD supports. - Sharing. Multiple hosts and devices touching one file raises consistency questions that a single-drive design must solve or sidestep. None of these kill the idea. All of them decide whether the idea leaves the lab. Notably, FADU scheduled the PCIe 6.0 host report ahead of partner samples, which gives the company a chance to publish hard numbers before outside engineers start probing. Five Predictions for FS-SSD Through 2027 These are our judgments, built on the milestones FADU stated. They are forecasts, not reported facts. 1. The Q1 2027 report becomes the credibility test. Write-amplification numbers will draw more scrutiny than the headline operations figure. A clean result would validate the F2FS-based design. A messy one would push the story toward a longer timeline. 2. At least one milestone slips. Hardware roadmaps that span Q1, Q3, and year-end 2027 rarely hold every date. The partner-sample window and the DPU reference design both depend on outside silicon and customer schedules. 3. Early adopters cluster around DPU-centric designs. The reported 800G E1 DPU reference design points to network-attached GPU clusters as the first fit, not general-purpose servers. 4. Rivals answer with their own offload pitches. Once a controller vendor shows a CPU-savings number, others usually follow with a counter-claim. Expect more talk of file-aware or object-aware drives during the next year. 5. The CPU shrinks in the data path but does not vanish. Even if FS-SSD hits its goals, servers keep CPUs for orchestration, preprocessing, and everything outside the storage path. The honest framing is less CPU per terabyte served, not zero CPU. For readers following how hardware shifts feed into model training and inference economics, our AI and machine learning coverage https://shattered.io/ai-machine-learning/ tracks the related infrastructure stories as they land. What Buyers and Engineers Should Do Now Nothing about FS-SSD requires action in 2026. Still, a few low-cost steps keep teams ready if the Q1 2027 results land well. First, measure your own baseline. Profile how much CPU your GPU nodes spend in the file system and block layer under realistic loads, including random reads through GPUDirect Storage if you use it. If your number is far below 30%, FS-SSD matters less to you. If it is near or above, you have a business case worth tracking. Second, ask your server and storage vendors what they know. FADU described OEM and ODM talks as early, so a vendor rep may have nothing yet, but asking now puts you on the list for the Q3 2027 sample window. Third, plan evaluation criteria in advance: write amplification, tail latency, recovery behavior, firmware update process, and security model. Having the checklist ready saves months once hardware arrives. Last, keep spending decisions separate from the announcement. No price exists in the public record, no commercial release has been confirmed, and no independent lab has validated the quoted performance. Treat FS-SSD as an option on the future, not a line item in this year’s budget. The Bottom Line on FADU’s File-Semantic SSD The takeaway: FADU has a clear thesis, a named architecture, and a staged plan to prove it, but the proof sits in 2027. The thesis is that host file-system and block-layer work, which FADU measured at about 30% of CPU cycles in GPUDirect Storage random reads, is worth moving into the drive. The architecture packs NFSv3 into NVMe commands and an F2FS-based layer into a sixth-generation controller. The plan runs from a PCIe 6.0 host report by Q1 2027 to partner samples in Q3 2027 and a DPU reference design at the end of 2027. The preliminary 460,000 operations per second at one core is an encouraging bench result, not a verdict. If the Q1 2027 write-amplification numbers hold up and partners can reproduce the gains, FS-SSD could change how AI storage nodes get built. If not, it joins the long list of smart-drive ideas that were right about the problem and early on the solution. FAQ: FADU FS-SSD and CPU-Free AI Storage What is the FADU FS-SSD? The File-Semantic SSD is a drive FADU introduced at the OCP Global Summit. Its sixth-generation controller runs firmware that handles file lookup, access checks, and file-to-address translation, so the host does not need to run those steps. Does FS-SSD really remove the CPU from AI storage? Not entirely, and not yet proven. FADU’s whitepaper describes a path with no host filesystem, no block layer, and no storage-server CPU. That is a design goal and a claimed access path. Servers still need CPUs for other jobs, and no independent test has confirmed the claim. How does FS-SSD talk to GPUs and DPUs? It carries NFSv3 requests inside standard NVMe commands. FADU says GPUs, DPUs, and network cards can reach files this way. The reports do not name which specific devices have been tested. What performance has FADU shown? A preliminary test reported 460,000 operations per second as the server configuration was reduced from 16 CPU cores to one. FADU also said the host file system and block layer used about 30% of CPU cycles during random reads through GPUDirect Storage. Both figures come from the vendor. When can customers get FS-SSD? Partner samples are expected in Q3 2027. FADU targets a report on FS-SSD testing in a PCIe 6.0 host by Q1 2027, and a reference design based on an 800G E1 DPU is targeted for the end of 2027. OEM and ODM discussions are at an early stage. How much will FS-SSD cost? No price has been identified in the available reporting. Anyone quoting a price before the sampling phase is guessing. What file system does FS-SSD use? The internal file layer is based on F2FS, a Linux file system designed for flash memory. FADU has promised hardware write-amplification figures in its Q1 2027 report, which will show how that design behaves on real flash.