M.2 AI Accelerator recommendations for 2026 / future roadmap?

[Oniom88]

Hi everyone,

I'm planning a new hardware deployment and could use some advice from the community and developers on which M.2 AI accelerators to look at for late 2026 and beyond.

The current hardware landscape for local object detection feels a bit stuck:

• Google Coral is outdated/EOL at this point and no longer recommended.
• Hailo-8 / 8L works great natively with Frigate, but the core architecture is getting pretty old.
• MemryX MX3 seems cool, but it's almost impossible to source or buy in Europe right now.

Given this, what are the current recommendations for building a future-proof setup today?

Also, I'm highly interested to know if there is a roadmap or any plans to support newer modules:

1. Is there a plan (or timeline) to add full native support for the Hailo-10H?
2. Are there any discussions around newer modules like the Radxa AICore DX-M1M, Axelera Metis, or UP AI Core XM?

Would love to hear what the community or devs are leaning towards these days for high-performance, next-gen setups.

Thanks!

[NickM-27]

If you're planning a new build why are you intending to use a m.2 based accelerator at all? Why not just get a system with an NPU and/or integrated GPU?

Most of the m.2 options only run a single instance of a single object detection model, making them scale poorly and not cover many available features.

[Oniom88]

Yes, it's a new build, but i'm trying do it cost efficiently, i purchased HP 840 G3 SFF i7 7700 with RTX 730 for 60$ and i will have 4-5x Dahua IPC-HDW2649TM-S-IL-0280B. I'm thinking to buy nvidia T400/600 or T1000 and ai hardware accelerator. please let me know what i can use - i want solution for few years use

[NickM-27]

I see, the problem with m.2 accelerators as I said is they are not very flexible. They are power efficient but the capability is limited in breadth. A GPU such as an Nvidia or something else like Intel Arc A370 or similar provides a similar (or slightly higher) price point, while being able to run all of Frigate's features at once.