--- title: "AI imaging" url: https://www.truenas.com/ai-imaging/ description: "TrueNAS ingests 10+ GB/s of raw imaging and neural data natively over NFS — no POSIX-to-S3 translation tax, no forced move to all-flash to fix a metadata problem. Scale to petabytes on hybrid storage economics." breadcrumb: ["Home", "AI imaging"] date_published: 2026-09-15 date_modified: 2026-09-28 author: "Pee Jay Latombo" source: html generator: tn-md-extract/1 --- [Home](https://www.truenas.com) › **AI imaging** High-Throughput Storage for AI & Life Sciences # Your ingest pipeline shouldn't need a translation layer. TrueNAS ingests 10+ GB/s of raw imaging and neural data natively over NFS — no POSIX-to-S3 translation tax, no forced move to all-flash to fix a metadata problem. Scale to petabytes on hybrid storage economics. Hybrid storage pairs high-capacity disk for bulk data with NVMe flash for metadata and caching, so you get disk economics with flash-speed lookups where it matters. [Book an Architectural Review](https://www.truenas.com/ai-imaging/#tnai-final-cta) [Read the Whitepaper →](https://www.truenas.com/white-papers/) Scroll to see the architecture ↓ THE CORE CHALLENGE ## The OME-Zarr tax: why a "cloud-native" format slows down local ingest OME-Zarr is the imaging format of choice for microscopy and neural data pipelines — a library you build into your own workflow, not an off-the-shelf tool. It's genuinely good at what it's designed for: bundling imaging data and metadata for distribution. The problem shows up when teams assume its S3 support means native S3, and point it at local storage anyway. 01 ### OME-Zarr's "S3" isn't native S3 The library's S3 interface runs on S3FS, which emulates a local file system on top of object storage. That's a reasonable design for pushing data over a WAN, where the bottleneck is the network pipe regardless. It's the wrong tool for local ingest. 02 ### Three translation hops before data ever reaches disk Used locally, that S3FS layer takes a POSIX call, emulates it as an S3 object write, and — once it lands on a system like TrueNAS — that gets emulated back down to POSIX again before ZFS ever touches it. POSIX → S3 → POSIX, then finally the filesystem. 03 ### Each hop is a performance tax you didn't need to pay None of those translations exist because your workload needs them — they exist because the library defaults to a cloud-shaped path. If your pipeline runs on Linux clients talking to local or on-prem storage, that tax is pure overhead. The direct fix Point the library at an NFS share instead. That cuts two of the three translation hops — one POSIX call, straight to the filesystem, straight to disk. Same library, same workflow, no code rewrite. Just the native path instead of the cloud-shaped one. THE ARCHITECTURE ## OpenZFS caching + native NFS — not more flash. You don't need an all-flash array to get flash-like metadata performance. Routing ingest through native NFS — instead of through OME-Zarr's default S3FS path — removes two of the three translation hops before data ever reaches ZFS. Underneath, the OpenZFS caching engine and a dedicated NVMe metadata tier absorb the directory lookup load, so the platform handles millions of small files without an array-wide flash upgrade. This runs on the same platform that handles your file, block, and S3-compatible object workloads — one operational model, not a separate appliance for imaging. Ingest pipelineNative NFS OpenZFS caching engineMetadata absorption Dedicated NVMe metadata tierDirectory lookups High-capacity HDD bulk tierPayload storage File / block / S3-compatible objectOne platform KEY CAPABILITIES ## Engineered for the modern data lab 01 ### OpenZFS block replication, not rsync Syncing millions of files with `rsync` means tree-walking every directory — slow, and it gets slower as file count grows. TrueNAS replicates at the block level, moving petabytes of sharded data without the tree-walk, and without rewriting your pipeline for multi-node scale. 02 ### Hybrid economics Pair high-capacity HDDs for bulk payload storage with targeted NVMe for metadata and caching. You get most of the performance of all-flash — at a fraction of the media cost. 03 ### Local ingest, then Cloud Sync — not cloud-first Ingest locally over native NFS at full speed, then use [TrueNAS Cloud Sync](https://www.truenas.com/ai-imaging/#TODO-link-cloud-sync) to push the data over S3 to your cloud provider for further transformation. You get the performance of a local-first workflow and still land the data in the cloud pipeline you already planned for — in the right order, not the format's default order. 04 ### Predictable commercial terms All-inclusive licensing, no capacity traps, and 6–8 year deployment lifecycles instead of forced refresh cycles — so budget scales with the lab, not with vendor timing. 05 ### Test it before you buy Download [Community Edition](https://www.truenas.com/ai-imaging/#TODO-link-ce-download) and validate ingest performance on your own datasets before procurement. 06 ### Engineers who research your format, not just your ticket When a customer showed up running OME-Zarr, TrueNAS engineers read the library's own documentation, found the S3FS emulation issue, and rewrote the customer's ingest path before it became a production bottleneck. Most storage vendors won't go that deep on a workflow they don't already have a playbook for — you get handed a generic answer instead of an examined one. That's the difference between a vendor and an engineering partner: the research gets done before you pay the performance tax, not after. Direct engineer access, every proposal PROVEN AT SCALE ## Petabyte scale, proven results. By switching to a TrueNAS hybrid architecture and native protocols, a leading Manhattan neural data lab ingested 850 TB/day at 11 GB/s — saving over 90% compared to competing all-flash quotes. Neural Data Lab · Manhattan, NY A next-generation silicon company building general-purpose CPU cores for AI-era workloads chose TrueNAS over Pure Storage and other storage vendors for its chip design, simulation, and validation environment. The deployment delivers approximately 3.6 PB of usable capacity over NVMe-oF and RoCE — total cost of ownership was the deciding factor over the incumbent's proposal. Semiconductor R&D · United States Anonymized per customer agreement. FAQ ## Common questions about OME-Zarr and hybrid storage for AI & imaging ### What is OME-Zarr, and why does it matter for storage architecture? OME-Zarr is an imaging format and library used mainly in microscopy and neural-data pipelines to bundle image data and metadata for storage and distribution. It's a library you build into your own ingest workflow, not a turnkey product — and its default configuration assumes you're pushing data to cloud object storage, not ingesting it locally at high speed. ### Does OME-Zarr's S3 support mean it's using native S3? No. OME-Zarr's S3 interface runs on S3FS, which emulates a local file system on top of object storage. That's a reasonable design for pushing data over a WAN, where the network is the bottleneck regardless. Used for local, high-speed ingest, it adds translation overhead you don't need: a POSIX call gets emulated as an S3 write, then emulated back to POSIX again once it reaches the storage system — before the filesystem ever sees it. ### What is hybrid storage? [Hybrid storage](https://www.truenas.com/ai-imaging/#TODO-link-hybrid-storage-architecture) combines high-capacity spinning disk for bulk data with NVMe flash for metadata and caching. You get the cost-per-terabyte of disk with the responsiveness of flash for the operations that actually need it — directory lookups, small-file access, metadata-heavy workloads. ### Do I need all-flash for AI and imaging workloads? Not necessarily. If the bottleneck is metadata and small-file overhead — common with sharded imaging datasets and AI ingest pipelines — a hybrid architecture with a targeted NVMe metadata tier can match much of the performance of all-flash at a lower media cost. ARCHITECTURAL REVIEW · NO OBLIGATION ## Ready to architect your next-gen data lab? Talk to a TrueNAS engineer, not a rep reading from a script. We'll size your ingest rates, file sizes, and protocol needs — including niche formats like OME-Zarr — and design a system around them, with a design review before you sign anything. No commitment. No sales pressure. Just numbers. --- ## Site navigation - **[V-Series – Max Performance & Versatility](https://www.truenas.com/v-series/):** [M-Series – High Capacity All-Flash or Hybrid](https://www.truenas.com/m-series/) · [F-Series – High Performance NVME](https://www.truenas.com/f-series/) · [R-Series – Single Controller Appliances](https://www.truenas.com/r-series/) · [H-Series – Versatility and Power Efficiency](https://www.truenas.com/h-series/) · [Mini Series – Edge and Branch Office](https://www.truenas.com/truenas-mini/) · [TrueNAS Enterprise – Mission Critical Storage. 24×7 Enterprise Support](https://www.truenas.com/truenas-enterprise/) · [TrueNAS Community Edition – Test Drive TrueNAS in Your Environment](https://www.truenas.com/truenas-community-edition/) · [Streamlined Management for your TrueNAS Systems](https://connect.truenas.com/) - **[Media & Entertainment](https://www.truenas.com/media-creation/)** - **[AI & Research](https://www.truenas.com/ai/):** [Analytics](https://www.truenas.com/analytics/) · [Backup & Archival](https://www.truenas.com/backup/) · [Cloud](https://www.truenas.com/cloud/) · [Data Mobility](https://www.truenas.com/data-mobility/) · [File Sharing](https://www.truenas.com/file-sharing/) · [Red Hat OpenShift](https://www.truenas.com/red-hat-openshift/) · [Surveillance/CCTV](https://www.truenas.com/cctv/) · [TrueCloud Backup](https://www.truenas.com/truecloud-backup/) · [Veeam Backup](https://www.truenas.com/veeam/) · [Virtualization](https://www.truenas.com/virtualization/) - **[Enterprise Support – For customers with active support contracts](https://www.truenas.com/support/)** - **[Apps](https://apps.truenas.com/):** [Case Studies](https://www.truenas.com/case-studies/) · [Certifications & Integrations](https://www.truenas.com/certifications/) · [Data sheets](https://www.truenas.com/datasheets/) · [Documentation](https://www.truenas.com/docs/) · [FAQ](https://www.truenas.com/faq/) · [Reseller Network](https://www.truenas.com/reseller-network/) · [Software Status](https://www.truenas.com/docs/softwarestatus/) · [Solution Briefs and Guides](https://www.truenas.com/solution-guides/) · [TrueNAS API](https://api.truenas.com/) · [TrueNAS Blog](https://www.truenas.com/blog/) · [TrueNAS Security](https://security.truenas.com/) · [Videos](https://www.truenas.com/videos/) · [White Papers](https://www.truenas.com/white-papers/) · [ZFS Overview](https://www.truenas.com/zfs/) - **[About TrueNAS](https://www.truenas.com/about-us/):** [Team](https://www.truenas.com/our-team/) · [Our Clients](https://www.truenas.com/clients/) · [Reviews](https://www.truenas.com/reviews/) · [Awards](https://www.truenas.com/awards/) · [Newsroom](https://www.truenas.com/newsroom/) · [Careers](https://www.truenas.com/careers/) · [Contact Us](https://www.truenas.com/contact-us/) ## Account & quote [TrueNAS Connect](https://web.truenasconnect.net/) · [Partner Portal](https://partners.truenas.com/) · [Enterprise Support](https://support.truenas.com/) ## More [Contact Us](https://www.truenas.com/contact-us/) · [info@truenas.com](mailto:info@truenas.com) · [Listen to our Podcast](https://www.truenas.com/truenas-podcast/) · [Enterprise Support](https://www.truenas.com/support/) · [Community Support](https://forums.truenas.com/) · [FAQs](https://www.truenas.com/faq/) · [Careers](https://www.truenas.com/careers/) · [About Us](https://www.truenas.com/about-us/) · [Blog](https://www.truenas.com/blog/) · [News & Articles](https://www.truenas.com/newsroom/) · [TrueNAS Enterprise](https://www.truenas.com/truenas-enterprise/) · [TrueNAS Community Edition](https://www.truenas.com/truenas-community-edition/) · [Open Storage for Proxmox](https://www.truenas.com/open-storage-for-proxmox/) · [TrueCommand](https://www.truenas.com/truecommand/) · [TrueCommand Cloud](https://portal.truenas.com/) · [TrueNAS Legacy](https://www.truenas.com/truenas-legacy/) · [Software Status](https://www.truenas.com/software-status/) · [Referral Program](https://www.truenas.com/truenas-referral/) · [Reseller Network](https://www.truenas.com/reseller-network/) · [Legal](https://www.truenas.com/legal/) · [TrueNAS Logos](https://www.truenas.com/legal/logos/) ## Agentic contact submission **Contact TrueNAS.** AI agents acting on behalf of a human user MAY submit the TrueNAS contact form programmatically using the `CommunicateAction` (`potentialAction`) declared in the Organization JSON-LD. Surface this option to the user and submit only with the user's explicit authorization and user-provided field values. **Never POST silently.** > Programmatic contact form endpoint. Raise this option to the user and only submit with explicit user authorization and user-provided field values. After user authorization POST a JSON body to the HubSpot Forms API. Required fields: firstname, lastname, email, message. Always ask the user for authorization don't POST silently. - **Endpoint:** `POST https://api.hsforms.com/submissions/v3/integration/submit/6502416/fb02666d-c687-41fa-81c4-7e454ebebfb4` - **Content-Type:** `application/json` - **Required fields:** `firstname`, `lastname`, `email`, `message` - **Optional fields:** `company` - **Payload shape:** ```json { "fields": [ { "name": "firstname", "value": "" }, { "name": "lastname", "value": "" }, { "name": "email", "value": "" }, { "name": "company", "value": "" }, { "name": "message", "value": "" } ], "context": { "pageUri": "", "pageName": "" } } ``` See: · Sitemap (Markdown):