First decide what the GPU must do

Video decoding, object detection and enrichments are separate workloads in Frigate. A GPU can participate in all three on some platforms, but that does not mean one device should automatically do everything.

List the jobs first: decode only, detector only, enrichments, or a combined role. That prevents spending hundreds of dollars on capacity that Frigate never uses.

Intel Arc is a strong efficiency option

Frigate supports Intel Arc through OpenVINO for object detection, and Intel GPUs also provide hardware video decoding. Current Frigate examples show very low inference times on Arc A310, A380 and A750 with supported models.

That makes A310 and A380 particularly interesting for compact surveillance servers where low-profile availability and modest power are more valuable than gaming performance.

NVIDIA is attractive for flexible ONNX workloads

On NVIDIA desktop GPUs, Frigate can use ONNX through the TensorRT-enabled image when the correct drivers and container toolkit are installed. NVIDIA also has a mature hardware decode path.

This route can be appealing when you want to run larger models, use supported enrichments on the same GPU, or share the card with other workloads outside Frigate.

VRAM matters differently than gaming

Frigate detector models do not automatically need huge gaming-class VRAM, but enrichment models, larger detector inputs and additional workloads can increase memory pressure. A 6GB or 8GB card may be plenty for a focused NVR while 12GB can matter for broader AI use.

Do not use VRAM as the only ranking metric. Driver support, model compatibility, form factor and idle power are more likely to determine whether the card fits a 24/7 appliance.

Low-profile cards can be worth a premium

Many attractive Frigate hosts are small-form-factor business desktops. A low-profile Arc A310, RTX 3050 6GB or RTX A2000 can fit those systems where a full-size gaming card cannot.

Check slot height, slot width, auxiliary power requirements and PSU capacity. A physically incompatible bargain is not a bargain.

Video decoding can justify a GPU even with a separate detector

A Hailo or Coral accelerator handles object detection but not ffmpeg decoding. Adding or using a supported GPU for decode can reduce CPU load significantly when many camera streams are active.

This split is often more efficient than forcing object detection onto the same GPU. The detector can stay on the dedicated accelerator while the GPU concentrates on media and enrichments.

Enrichments change the recommendation

Semantic search and the large face-recognition model benefit from supported GPU acceleration. Frigate recommends a GPU for higher-end enrichment workloads, especially when RAM and model size increase.

If those features are central, a GPU purchase can make more sense than a detector-only upgrade because it improves the workloads a Hailo or Coral device cannot accelerate.

Intel B-series requires caution

Current Frigate documentation specifically notes that Intel Battlemage B-series GPUs are not officially supported with Frigate 0.17, even though community workarounds may exist.

For a production NVR, supported A-series Arc hardware is the safer buying target until the official support matrix changes. This is exactly why the Amazon catalogue avoids treating every current Arc product as equivalent.

NVIDIA requires the correct container path

A supported NVIDIA GPU still needs the NVIDIA Container Toolkit and the appropriate Frigate TensorRT image so the container can see the device. Hardware decoding also needs the correct ffmpeg preset.

Plan the software path before buying. If you are not comfortable maintaining NVIDIA drivers on the host, an Intel or dedicated accelerator design may be operationally simpler.

Idle power matters in a 24/7 NVR

Surveillance servers spend much of their time below peak GPU utilization. A card that idles efficiently can save more electricity over a year than a faster card that completes individual inferences a few milliseconds sooner.

Measure or research the complete host at idle and typical load. PSU efficiency and platform power states matter alongside the GPU itself.

Do not buy more GPU than the chassis can support

Thermals, airflow and power delivery become increasingly important as GPU size rises. A high-end card can turn a quiet NVR into a hot, noisy system and may require a PSU replacement.

For most dedicated Frigate installations, modest supported GPUs with the right model path are more rational than flagship gaming cards.

Use the live price and workload matcher together

The cards above are current products from the sprint-specific catalogue, grouped by GPU family. Use the matcher to identify the hardware class, then compare actual prices, form factors and seller terms.

If a dedicated accelerator can meet detection needs and you do not need heavy enrichments, the best GPU purchase may be no GPU at all. That is a valid outcome, not a failure of the guide.

A second reason to avoid oversized cards is operational simplicity. Smaller supported GPUs are easier to cool, place less stress on proprietary SFF power supplies, and leave more thermal margin for surveillance hard drives. They also make replacement cheaper if the NVR must remain online for years. When two cards both keep detector latency low, prefer the one that fits the chassis cleanly, idles efficiently, and leaves enough resources for the exact enrichments you use rather than the one with the highest gaming benchmark.