NVIDIA uses a distinct Frigate software path

Frigate uses the TensorRT-tagged image for supported NVIDIA GPU acceleration and can run ONNX detector models through that backend. The host also needs the NVIDIA Container Toolkit and compatible drivers.

That stack is mature but adds maintenance compared with an all-Intel default image. Decide whether the flexibility is worth the extra software layer in your environment.

RTX 3050 6GB fits many compact builds

The 6GB RTX 3050 is available in low-profile variants that can turn an SFF business desktop into a capable GPU-backed NVR. It provides enough memory for many Frigate tasks without requiring a large gaming chassis.

Check whether the exact card draws power only from the slot or requires an auxiliary connector. That detail can decide whether a refurbished office PC is suitable.

RTX A2000 is built for small workstations

RTX A2000 is a professional low-profile card designed for compact workstations. Its form factor and efficiency can be more important to an NVR builder than gaming benchmark position.

Used and refurbished pricing varies widely, so compare condition and seller warranty carefully. A low advertised price can disappear once risk and missing accessories are considered.

RTX 4060 is for broader GPU use

RTX 4060 provides more modern GPU headroom and can make sense when Frigate shares the card with other CUDA or AI workloads. It may also provide more margin for heavy enrichments than a small low-profile card.

The tradeoff is physical size and system design. Many 4060 boards require a full-height case and a more capable PSU.

RTX 3060 remains a memory-oriented value option

RTX 3060 12GB can still be attractive when VRAM capacity matters and current pricing is favorable. For a pure Frigate detector, however, that memory may be more than you need.

Do not buy old hardware solely because it has more VRAM. Driver support, efficiency, seller condition and form factor remain important.

Hardware decoding is a separate reason to choose NVIDIA

Frigate supports NVIDIA hardware video decoding with the appropriate ffmpeg preset. On camera-heavy installations, decode acceleration can save substantial CPU time even if object detection runs on another device.

This makes a hybrid Hailo-plus-NVIDIA architecture legitimate: Hailo handles detector work while NVIDIA handles decoding and enrichments.

Enrichments can use the same NVIDIA card

Frigate automatically detects supported NVIDIA GPUs for enrichments when the correct TensorRT image is used. Semantic search and large face recognition are examples of workloads that can benefit from GPU acceleration.

If enrichments are a priority, GPU selection should include their memory and compute requirements rather than focusing only on detector latency.

A Coral or Hailo does not make NVIDIA redundant

Dedicated detectors are efficient at object detection, but they do not replace a general GPU for every Frigate AI feature. A smaller NVIDIA card can therefore complement rather than compete with a Hailo module.

The right combination depends on chassis slots and power. Sometimes one GPU doing several jobs is cleaner; sometimes separating the workloads is easier to tune.

Container setup must be part of the purchase checklist

The NVIDIA Container Toolkit, Docker GPU reservation and correct Frigate image are prerequisites, not optional tuning. Plan those steps before the card arrives.

If your platform or hypervisor makes GPU passthrough difficult, an Intel iGPU or dedicated accelerator may produce a simpler system even if NVIDIA has more raw compute.

Power and fan behavior affect 24/7 use

A surveillance server sits in a different environment from a gaming PC. Fan-stop behavior, idle power, thermal exhaust and acoustic profile matter because the card may run continuously in a small chassis.

A well-cooled low-profile card can be a better NVR component than a faster model that constantly heats the drive cage.

Also consider where the heat is exhausted. In compact cases, a blower-style or carefully ducted card can protect nearby storage better than an open-air cooler that recirculates warm air around drive bays. Monitor drive and GPU temperatures together after deployment.

Model selection matters more than brand loyalty

Frigate supports multiple detector backends and model types. Pick the model family and required Docker image first, then choose the NVIDIA card that meets the workload with reasonable headroom.

This avoids buying a card for a benchmark that has little relationship to the detector you will actually deploy.

Practical choice

Choose RTX 3050 6GB or A2000 for compact, efficiency-oriented systems; choose RTX 4060 when you need more general GPU headroom; consider RTX 3060 12GB when memory and price create a strong value case.

The matcher above converts camera activity, enrichments, chassis and power limits into a tier before you compare the live listings.

Before finalizing an NVIDIA card, test the entire software chain on the target host: the proprietary driver must load correctly, Docker must expose the GPU, the TensorRT Frigate image must start with the expected providers, ffmpeg should show hardware decoding, and the chosen ONNX model should run without falling back to CPU. This validation matters more than a small benchmark advantage because an NVR is an appliance. A slightly slower card with a clean driver path and predictable thermals is usually the better long-term choice.