[Detector Support]: Should I migrate to Nvidia GPU?

[TheKalpit]

Describe the problem you are having

I'm running a 12 camera setup on a single coral TPU for over 2 years now. The detector inference is typically ~8ms. The detection accuracy, I believe, is good enough. I'd be happy if it can be improved somehow... that's what my question really is.

I also have a Nvidia GTX 1050 Ti GPU lying around. I configured frigate on it, to see if it performs better. I've been running it with 2 cameras only for about 1 day now, and the inference speed is ~23ms. I'm using yolo nas (small) model.

My goal is to improve accuracy, and maybe try out the new face recognition features in v0.16.

1. Is this inference speed expected or can it be improved by config changes?
2. Should 1050 Ti perform better (inference speed & accuracy) than the coral TPU?
3. Is the 1050 Ti acceptable for the new face recognition feature?

Thanks!

Version

0.16.0-c2f8de9

Frigate config file

# nvidia config (relevant parts only)

detectors:
  onnx:
    type: onnx

model:
  model_type: yolonas
  width: 320
  height: 320
  input_pixel_format: bgr
  input_tensor: nchw
  path: /my-models/yolo_nas_s.onnx
  labelmap_path: /labelmap/coco-80.txt

ffmpeg:
  hwaccel_args: preset-nvidia

docker-compose file or Docker CLI command

-

Relevant Frigate log output

2025-08-23 20:13:43.367663394  [INFO] Preparing Frigate...
2025-08-23 20:13:44.674723220  [INFO] Starting Frigate...
2025-08-23 20:13:47.375880127  [2025-08-24 01:43:47] frigate.util.config            INFO    : Checking if frigate config needs migration...
2025-08-23 20:13:47.400016447  [2025-08-24 01:43:47] frigate.util.config            INFO    : frigate config does not need migration...
2025-08-23 20:13:47.522221982  [2025-08-24 01:43:47] frigate.config.config          WARNING : Frigate+ is configured but clean snapshots are not enabled, submissions to Frigate+ will not be possible./
2025-08-23 20:13:47.524251744  [2025-08-24 01:43:47] frigate.app                    INFO    : Starting Frigate (0.16.0-c2f8de9)
2025-08-23 20:13:47.534707032  [2025-08-24 01:43:47] peewee_migrate.logs            INFO    : Starting migrations
2025-08-23 20:13:47.535341067  [2025-08-24 01:43:47] peewee_migrate.logs            INFO    : There is nothing to migrate
2025-08-23 20:13:47.549534054  [2025-08-24 01:43:47] frigate.app                    INFO    : Recording process started: 323
2025-08-23 20:13:47.550081086  [2025-08-24 01:43:47] frigate.app                    INFO    : Review process started: 328
2025-08-23 20:13:47.553183049  [2025-08-24 01:43:47] frigate.app                    INFO    : go2rtc process pid: 138
2025-08-23 20:13:47.567582831  [2025-08-24 01:43:47] detector.onnx                  INFO    : Starting detection process: 348
2025-08-23 20:13:47.568947672  [2025-08-24 01:43:47] frigate.detectors.plugins.onnx INFO    : ONNX: loaded onnxruntime module
2025-08-23 20:13:47.570148718  [2025-08-24 01:43:47] frigate.detectors.plugins.onnx INFO    : ONNX: loading /my-models/yolo_nas_s.onnx
2025-08-23 20:13:47.596092534  [2025-08-24 01:43:47] frigate.app                    INFO    : Output process started: 367
2025-08-23 20:13:47.674326153  [2025-08-24 01:43:47] frigate.app                    INFO    : Camera processor started for gate_48: 390
2025-08-23 20:13:47.691729383  [2025-08-24 01:43:47] frigate.app                    INFO    : Camera processor started for gate_49: 392
2025-08-23 20:13:47.776089230  [2025-08-24 01:43:47] frigate.app                    INFO    : Capture process started for gate_48: 412
2025-08-23 20:13:47.858349566  [2025-08-24 01:43:47] frigate.app                    INFO    : Capture process started for gate_49: 423
2025-08-23 20:13:48.506437632  [2025-08-24 01:43:48] frigate.detectors.plugins.onnx INFO    : ONNX: /my-models/yolo_nas_s.onnx loaded

Install method

Docker CLI

Object Detector

TensorRT

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

1. Inference Speed
   An inference speed of ~23ms for YOLO-NAS (small) on a GTX 1050 Ti is in line with documented expectations for similar Nvidia GPUs and YOLO-NAS models(1)(2). For comparison, the Coral TPU typically achieves ~8-10ms inference using the default SSDLite/MobileNet or similar models(1)(2). YOLO-NAS models are generally more accurate but computationally heavier than the default Coral models, so the slower inference is expected(1)(2)(3). You can experiment with smaller YOLO or MobileNet models to improve speed, but this may reduce accuracy.

2. Performance: GTX 1050 Ti vs Coral TPU
   The Coral TPU is optimized for speed with MobileNet/SSDLite models and achieves very low inference times(1)(2). The GTX 1050 Ti can run more accurate and complex models (like YOLO-NAS), but at a slower speed, especially for entry-level GPUs(1)(2)(3). For raw speed, Coral is hard to beat; for accuracy, Nvidia GPUs with modern models may give you an edge, but at the cost of higher inference times(1)(2)(3).

3. Face Recognition on GTX 1050 Ti
   The "large" face recognition model in Frigate is designed to leverage an integrated or discrete GPU for improved accuracy(4)(5). The GTX 1050 Ti is supported for GPU-accelerated enrichments, including face recognition, when using the appropriate Docker image (e.g., -tensorrt)(5). This means your card is suitable for running the new face recognition features with the "large" model(4)(5).

Summary:

• Your inference speeds are expected for the hardware and model chosen(1)(2)(3).
• The Coral TPU is faster for lightweight models, but Nvidia GPUs support more accurate, heavier models at the expense of speed(1)(2)(3).
• The GTX 1050 Ti is acceptable for GPU-accelerated face recognition with the "large" model(4)(5).

📚 Sources:

1. Recommended hardware
2. Supported Hardware
3. Supported Hardware
4. Face Recognition
5. Enrichments

---

Have another question? Just tag @inkeep.

[dbro]

@TheKalpit You can try running a YOLO v9 "s" 320x320 model on the Coral by following the (new!) instructions here
https://github.com/dbro/frigate-detector-edgetpu-yolo9
Detection latency is 11ms (3rd gen i7 CPU with a mini-pci-e coral device)