Hacker News Reader

Comment by elitistphoenix | Hacker News Reader

Google Coral Accelerator is basically abandoned these days though

Luckily Frigate works with a ton of different accelerators, like the Hailo, Intel's iGPU, even some Arm GPUs now too.

I have it running on an Orange Pi 5 with the Rockchip NPU, very impressed with that being supported and working so well for object detection.

single camera?

Two cameras, CPU usage is very low (NPU spikes when there’s movement). I suspect I could add more without too many problems.

wow.. thats sounds very promising.

Can you recommend a quality online community that do the same thing that I could lurk in for while to soak up some knowledge??

The documentation is rather scarse on performance numbers, but it looks like the hierachy of price/performance is like Intel iGPU ("free"), Intel A310, Nvidia GPU.

I'm explicitly leaving out the Coral TPU, since it's been reported that the newer Intel CPUs (Core Ultra) seem to provide the same performance with it's iGPU.

Still works with frigate, although I've heard that modern (whatever that means) CPUs can do as good a job as the Coral TPU, making it somewhat redundant.

I ain't running it on a modern CPU though, so I'm happy with the Coral.

Anecdata: i5-6500 did recognition in about 15ms, Coral TPU (M2 variant) does it in about 7.5ms - so… probably could’ve done without it in hindsight…

Frigate's docs has some detection speeds listed: https://docs.frigate.video/frigate/hardware/#openvino .

I believe it also has an advantage of being able to run bigger models like YOLO-NAS. Going off Frigate+ documentation: https://docs.frigate.video/plus/#supported-detector-types

Reply on news.ycombinator.com