About and methodology

Where every number comes from, and what we do to it

DetectFPS is a community benchmark database for people running AI security cameras locally with Frigate NVR. It collects the inference speeds, camera counts and power figures people post about their own hardware, keeps a link to every source, and turns them into medians, a capacity calculator and side-by-side comparisons. It is independent and not affiliated with the Frigate project.

Where the numbers come from

Public postsGitHub discussions, Reddit threads and blog posts where someone shared their Frigate numbers. We copy the figures as written, never re-test them, and link the post. If the source gave a range, the range is stored; if it only gave an upper bound, that is stored as “under N ms”.

Reader submissionsThe submit form asks for the inference time from Frigate's System page, the model, camera count and version. Each submission is read by a person, checked for obvious mistakes, and added to the dataset by hand. Nothing is published automatically.

Frigate docs referencesThe Frigate project's own hardware page lists approximate inference times for many devices. Those rows are included so the database is useful from day one, but they are always marked Frigate docs, kept out of “Latest builds”, and never described as user builds.

As of 2026-10-19 the database holds 173 builds across 69 devices: 57 from the community and 116 docs references.

How numbers are versioned

Every build records the Frigate version exactly as the source wrote it, shown as a badge like 0.17.2. The newest release (0.18.2 today) is set in one place, site.json, and a community build gets the older version, may be outdated label when it is more than two minor versions behind: with 0.18 newest, that means 0.15 and older. Builds that never stated a version get no label rather than a guess. Dates are stored as the source wrote them, with their precision, so a post dated only “Nov 2025” is never shown as a specific day.

How numbers are flagged

problem reportThe thread was a bug report; the number is what the user saw while something was wrong.

misconfigurationWrong model for the hardware, or a similar setup mistake. Usually excluded from stats.

anomalyFar outside what the hardware normally does, cause unknown.

known issueA documented bug affected the setup, but the reading itself is valid and still counts.

Flagged builds stay visible in every table with their reason, so nothing is silently dropped. Each build also carries a yes/no “usable for stats” field; builds marked no are left out of every median, leaderboard and calculator result. Today that is 5 of 173 builds.

Why docs numbers are labelled separately

The project's reference figures come from the maintainers' own test setups, usually with no camera count, no version and no power draw, and they are rounded. They are good for ranking devices against each other, and poor for predicting your exact install. So they count toward medians, but every place they appear says so, and a toggle on the home page lets you rank community builds only.

How the calculator works

capacity = detectors × 1000 ÷ median inference ms
demand = cameras × detect fps × activity (Quiet 0.25 · Typical 0.5 · Busy 1.0)
headroom = demand ÷ capacity → comfortable ≤ 60% · tight 60–100% · overloaded > 100%

The median is taken over the usable builds that match your device and model. MemryX devices report throughput rather than latency, so their median total fps is used instead of 1000 ÷ ms. The result lists every build behind the estimate and says how many there are, with a low-confidence note under three. “Cameras supported” on hardware and compare pages uses the same math at 5 fps and Typical activity: the number of cameras that stays under 60% of capacity. It is an estimate: resolution, motion masks, zones and the host CPU all change the real number.

Prices and affiliate links

A price appears only after it has been checked by hand on a real product page, with the date it was checked. Everything else says “price not tracked yet”. The speed-vs-price chart and the best-value ranking appear once at least three devices have a verified price. Price changes are covered in the weekly newsletter rather than by per-device alerts, because the site has no backend that could send them.

DetectFPS may earn a commission when you buy through links on this site. It never changes which builds we show or how they rank. Buy buttons are labelled “affiliate link” whenever one is used.

Download the dataset

builds.json173 buildsEvery field described on this page, one object per build, with source links.Download builds.json

hardware.json69 devicesNames, aliases, detector types, guide links, verified prices with dates.Download hardware.json

Updated 2026-10-19. Licence: CC BY 4.0, which means you can use and share the data as long as you link back. A one-line credit is enough: Data from DetectFPS (detectfps.com), CC BY 4.0.

Corrections and contact

Spotted a wrong number, a dead link or a build that should be flagged? Send the build id and what's wrong to hello@detectfps.com, or submit the corrected numbers through the form with a note. Corrections are applied by hand, usually within a week.

Not affiliated with Frigate

DetectFPS is an independent community project. It is not affiliated with, endorsed by, or maintained by the Frigate project or its contributors. Frigate is the name of their software; it appears on this site only to say which software the numbers were measured with.