seeed studio Coral M.2 Accelerator B+M Key

Fits right into an m.2 slot and out of the way. Runs as it should and the documentation is pretty extensive.

For what it is meant for it is very impressive. I set mine up currently for low-res image detection with frigate running in a Docker container via an Arch Linux host. It performs detections in I believe 6ms or so for me.

Purchased in order to speed up object filtering on my Frigate (Truenas EE) install. For reference the mother board is a Gigabyte Z590. After figuring out I needed to upgrade my processor to support the second M.2, and installing drivers and such the install went flawlessly. The system performance is night and day (three IP cams). I dropped one star since it only came in antistatic packaging, not instructions.

I was looking for alternatives for hosting my own cameras with an AI for object detection. I wanted a small form factor PC instead of having to rely on old gaming PCs with a graphics card to heat up my closet. I ran across this, and decided to try it in an HP Elite Desk Mini 800 G5. It has two NVME slots as well as the WiFi/Bluetooth slot. I had hoped to use the other Coral M.2 with two TPUs, but my WiFi slot wasn't fast enough. I went with this version that fits in my second NVME slot. It works great with CodeProject AI. CPU detection without this was over 400ms, added this and it was comparable to my full desktop with a GTX 970 at less than 60ms detection time. Sometimes even less than 20ms.

I was very pleased by this module as it enables computer vision workflows, but do not expect to be running LLMs on it yet.Keep in mind you will not be able to mount this and still use a regular case.On a side note, I wish the community around it was stronger and more inventive, as it feels like an abandoned project now.

Using in my SFF computer for Frigate. Works quite well on Unraid under Docker.

I added this card to my frigate server as the CPU was not able to handle object detection very well. After adding this card and getting things setup properly, my frigate is much more stable and the CPU hardly takes a hit.

I've been using the USB version of the Coral TPU to aide my Frigate NVR setup in object detection - which is a great way to gain access to the magic of TensorFlow Lite. The version we're looking at here has essentially the same features and same performance as the USB version, but doesn't depend on a USB 3.0 port (which is likely a plus, if you run it from a potentially crowded/packed space, potentially in a server rack). Being that the M.2 version runs from PCIe, it could see a *minimal* performance bump from reduced latency compared to USB, but I would expect this to be negligible.The real magic here is what the device is able to do. While you can absolutely run a sizeable security camera setup through Frigate NVR without a TPU, there are some very real, very tangible upsides to adding this (rather cheap) device into your setup. Frigate's object detection functionality becomes an order of magnitude faster, while simultaneously significantly dropping various other resources consumed. I've been running mine on an older generation i7-based. off-lease MFF HP desktop mini setup with 16GB of RAM, and no discrete GPU. This currently monitors just over a dozen cameras (all 4K or better), and is able to monitor all streams in real time, with inference time for all cameras combined averaging 10-11ms at all times. Can't beat that for $35!The only downside I can think of - for this particular version - is that there have been some recent rumblings about the PCIe version slowly starting to lose support from the Google side of things, with fewer and fewer updates coming through, and bugs going un-addressed longer and longer. Fortunately there are several ways people in the community have identified so far, that will allow you to keep using these and build your own supporting kernel versions etc. so at least there is that.Regardless, definitely get you one of these if you are running Frigate and don't have a TPU in the mix already!