Hardware programme

Lend us hardware. We publish what we measure.

OpenNVR runs on hardware people already own, and the only way to say "it runs on this board" honestly is to run it there. If you make, sell or simply have a board, an accelerator, a camera or a mini PC we have not measured, lend it to us or give us remote access, and we will put the whole stack on it, measure every role with the public protocol, and publish the results with your name on them.

How it has worked so far

In October 2026 Indus Technologies put an Arduino Ventuno Q on our network over a VPN. Two weeks later the Tier-0 detector, the captioner, the camera agent, speech to text, search and plate reading all ran on its Qualcomm Hexagon NPU, each number measured against the same board's CPU and published: the benchmark post, nine reports in the repository's bench folder, a row in the support matrix, and a -qualcomm flavour of every image that the installer now picks by itself. The board never left their building. That is the programme.

What we do with your hardware

  • Port, if needed. If the accelerator has an ONNX Runtime provider, a vendor runtime or a compiled-model zoo, we build the image flavour for it and wire it into OPENNVR_ACCEL=auto, so your customers get it with one setting, not a wiki page.
  • Measure every role, not one model: the always-on detector under a growing camera count, the ONNX adapters, captions and visual questions, the agent's language model, speech to text, search embeddings, plate reading, video decode, and the whole stack under a real camera for at least half an hour. CPU baseline on the same machine, always.
  • Publish all of it: a report per role in the repository with the commands that produced it, a row in docs/HARDWARE.md, and a blog post when the results are worth one. Bad numbers are published too; a board that decodes in software or an NPU that rejects a model is written down as such, with the reason.
  • File what we find. Driver bugs, kernel issues and runtime traps go upstream or to your engineers with logs, as the Ventuno Q's video decoder issue did.

What we need from you

NeedDetail
The hardware, one of two waysRemote access: SSH over a VPN of your choice (Tailscale worked well) to the board on your bench, with Docker installed and a sudo account. Or a loan: ship the board; we return it or keep it, as you prefer, and say which in the report.
TimeTwo to four weeks of access. A port with a new runtime takes the first week; the measurements the rest.
An engineer we can askOne contact who knows the platform. Most of a port is finding which device nodes, libraries and firmware a container has to borrow from the host, and one answer from the right person saves a day.
Documentation and runtimesThe SDK, the runtime libraries and their licence terms, and the model zoo if there is one. If the runtime cannot be redistributed we download it at first start, as we do for Qualcomm's.
Permission to publishYour company's name, the board's name, and the numbers. We do not publish pre-release products until you say so.

The rules we hold ourselves to

  • Numbers are as measured. No figure is ever taken from a datasheet and presented as ours, and no result is omitted because it is unflattering. You see the draft before it goes out and can correct facts about your product; you do not edit the measurements or the verdict.
  • Lending hardware buys the measurement, not the result. Sponsorship and loans are acknowledged in the report; they change nothing in it.
  • Your bench stays yours. On a remote board we run only OpenNVR, its fake-camera rig and public test clips. No customer footage, no credentials of ours left behind, and we tell you what we installed so you can wipe it.
  • Everything we build is open. The flavour, the adapter changes and the benchmark tools go into the public repositories under their existing licences.

What you get

  • A supported, measured path for OpenNVR on your hardware, selectable from one setting, in the next release.
  • Your name and the board's name in the report, the support matrix, the blog post and on our sponsors page, for as long as the numbers stand.
  • Independent, reproducible numbers your own customers can rerun, with the commands that produced them.
  • Bug reports with logs for whatever we hit on the way, and a team that has already been through this on three architectures to compare notes with.

What we would most like to measure next

HardwareWhy
NVIDIA Jetson Orin Nano Super, Orin NX, Thorthe nvidia flavour and the comparison people ask for most
Raspberry Pi 5 with the AI HAT+ (Hailo-8L, Hailo-8), Hailo-10Hthe small-site board; a Hailo path for the detector and the vision adapters
Rockchip RK3588 boards (Orange Pi 5, Radxa Rock 5)RKNN NPU plus hardware decode on a board many people already own
Other Qualcomm Dragonwing boards (IQ9, QCS6490, Rubik Pi)whether the qualcomm flavour carries over and which Gemma bundle each needs
Intel Core Ultra 200V mini PCs, AMD Ryzen AI mini PCsthe 48 TOPS NPU class, and a first ROCm / Ryzen AI measurement
Industrial IP cameras with on-camera analytics, PTZ and thermal camerasONVIF profiles, substreams and events we have not seen in the wild

Anything not on the list is welcome too. If people run OpenNVR on it, we want to know how it behaves.

Sponsoring the work

Hardware is one way to sponsor OpenNVR; funding is the other. The project is free software maintained by a small team, and a sponsor can fund a named piece of work: a hardware flavour, hardware video decode on a platform, a new adapter on an accelerator, a benchmark round, or a month of the camera agent's development. Sponsored work is scoped in public, delivered in public, and credited on the sponsors page and in the release notes. It does not buy a verdict, a hidden result or a closed feature. If you would rather sponsor the project as a whole, that is welcome too; tell us what you want it to go towards and we will say what it buys.

Offer hardware or sponsorship

Write to contact@opennvr.org with the board or accelerator, whether it is a loan or remote access, the window you can offer, and who we may ask questions. Or open a hardware offer on GitHub, which asks the same questions in public. We answer within a week and say plainly whether we can take it on and when.