It's taken me time to come to grips with this one. A 2017-era low end Cortex-A cpu with just enough RAM to survive & a GPU that seems most likely to never ever ever be in any way supported seems like a hard sell. That GPU is worse than worthless right now & has been since before 2017. At least the pros at Bay Libre are doing a decent job upstreaming good support for a lot of the rest of the MT8167 cpu, so although the PowerVR GPU seems like it will likely never ever be good for anything the rest could be fairly capable, if on the small side.

Once I got over some of the initial disappointment in the low-end CPU with agonizingly painfully unusable GPU though, & reflected a bit, I felt better. Oh sure, Coral is only $35 for the m.2 edition. And that made the board feel expensive. But the Coral USB edition is $59, and frankly, that's all most single-board-computers (SBCs) will be able to run. Alas most single-board-computers don't have a m.2 slot, and it's kind of bulky when they do. This thing is small. And, if you look at it from the $99 cost - $59 usb accelerator perspective, this is a modest but very reasonable $40 MT8167 SBC, in a small form factor, with a super neat & ultra-promising mega-efficient inference engine on it. It's taken me time, I felt cheated at first, but now, this board seems really neat. The SoCs onboard connectivity is decent (802.11ac and BT5). Both good & competent, but also lacking modern MU-MIMO which we all desperately hope will radically radically improve our AP's utilization. So, again, connectivity a bit long in the tooth, but also at least somewhat competent. Overall, still a very-slightly-overpriced-for-it's-age $40 TPU mated closely to a perhaps-maybe-fair $60 price for the recent-ish mini-TPU. An interesting board!

Needs benchmarking to see if slow 4x A35 and 2 GiB of RAM are good enough for that TPU.

Jetson Nano, and Rockchip RK3399Pro and RK3588 devices might be better and cheaper.