MX3 M.2 AI Accelerator

I bought this to install in a Raspberry Pi 5 setup installed in an Argon ONE v5 case with dual M.2 slots. I have a NVME drive in one slot and I wanted to install the MX3 in the ither slot to support detection in Frigate NVR. The problem I ran into was the MX3 gets quite hot running continuous, especially nest to the NVME drive. After following the installation instructions from Frigate, it worked fine, though slower than expected. Biggest issue you will have is getting rid of the heat if pairing it with a RPI5.

What makes this product stand out from the other accelerators is how everything seems to be open source. They have what they call a developer hub which contains everything you need to get started EVEN if you don’t know much about AI, just use their awesomely detailed tutorials or run their publicly available examples. And here’s the kicker, other accelerator manufacturers ask you to pick a model from their heavily tuned collection, this MemryX one doesn’t need that and actually they encourage you to use your own model as is. Honestly, I have been having a blast with this thing, I haven’t enjoyed working on something this much in a long time. kudos to the developers behind this.

Tried using this product with Frigate as they had a sale for $50. The guide seemed straightforward, but I could not get this device to work at all. Mind you, I used Proxmox and passed through the device to a Ubuntu 24.10 VM. Your mileage may vary. I was able to install the drivers, but their management service would start and hang. Would not ever identify the device. Frigate would through crazy errors. The documentation has absolutely no troubleshooting guidance whatsoever and the documentation from Frigate for this device seems to be written entirely by the makers as well with no troubleshooting guidance. I'd stay far away from this device unless you are running it bare metal on a fresh install.

I bought this for local AI and hardware experimentation, and it has been a strong little accelerator for that kind of work. I ended up building some Windows-side tooling around it, and it has been useful for validation, telemetry, and testing real local workflows. Best fit for developers and tinkerers who want to work directly with the hardware and software stack, not people expecting a completely plug-and-play consumer device. See my GitHub for a windows first app.

I love this thing

This thing is cool

As stated by others if you want to use this in Proxmox in your homelab with frigate this accelerator, it's driver and tools are not ready for primetime. Most complain about Coral TPU but I have been able to set up my Coral TPU and run successfully. Maybe it is just a matter of time before it works as I want but for now it's a paperweight.

Pretty awesome little accelerator! I'm eager to keep working on development with this and seeing where it goes.