16 to 32 cameras with full AI — on an €80 used graphics card

Most "AI NVR" systems have an awkward sentence buried in their setup guide: buy this accelerator first. A Coral stick, a specific new GPU, some dongle. For a home or small-business camera system, that's backwards — you shouldn't have to buy hardware before you can find out whether the software is any good.

Calyston does it the other way around. The AI runs on the CPU you already have, on day one, with nothing extra to buy. That's enough for a handful of cameras with motion-triggered detection. And when you want more — more cameras, or the heavier and more accurate detection model — you don't need a special box. Any NVIDIA graphics card does it, including a cheap used one.

We didn't want to claim that without proving it, so we tested on deliberately humble hardware: a second-hand GTX 1060, the kind you can find for around €80. Here's what we measured.

The numbers (measured, not estimated)

The accurate detection model (YOLOX-L) on the CPU of a mid-range machine takes about 477 ms per frame. On that same machine, with the €80 card doing the work:

That's an 8.8× speedup — and it changes the whole picture. At ~54 ms the card can run roughly 18 detections every second. Because Calyston only runs the AI when something actually moves (not on every frame of every camera), that comfortably covers 16 to 32 cameras with the accurate model, not the lightweight one.

And it's not only object detection. Calyston's natural-language search — the feature that lets you type "person in a red shirt near the gate" and jump straight to the clip — runs on the very same card. Both the detector and the search model together use about 600 MB of the card's memory, so even a 6 GB budget card has room to spare.

One thing worth stating plainly: the GPU doesn't change what gets detected, only how fast. The results are identical to the CPU — we ran the same frames through both and got the exact same detections. The card buys you scale and speed, never different answers.

Why we support old cards on purpose

Here's the part most software won't tell you: the newest AI libraries have quietly dropped support for older GPUs like the GTX 10-series. The official "just install it" path simply won't run on them anymore. We did the unglamorous work to keep those cards working, because for a budget camera build, a €80 used card is the smartest money you can spend — and it shouldn't be locked out by a version number.

So if you've got an old gaming GPU gathering dust, or you can grab one cheap, Calyston will use it.

The honest catch

We're not going to call it free of trade-offs. Turning on GPU acceleration pulls a one-time download of a few gigabytes (the GPU runtime), and you need NVIDIA's driver installed. That's it — no account, no cloud, no per-camera fee. The CPU mode stays the lean default; the GPU is there when you want to scale up, and it's included on every plan. We're not going to charge you to use a graphics card you already own.

The bottom line

Start on the CPU you have. If you outgrow it, a cheap used GPU takes you to 16–32 cameras with the accurate model and natural-language search — measured, not promised. No special hardware, no vendor lock-in, no surprise.