However, if you got a cluster of them working together as a phased array - then you not only overcome the limitations, but surpass alternative offerings. But is the tech ready for that, maybe not and who knows - inter Cubesat communication may end up being a case of using lasers for that short few meter formation cluster. Certainly some interesting challenges and yet such things are within reach.

You can’t maintain a cluster of things in orbit without actively expending propellent. At best you can have either a long line of cube sats in the exact same orbit, or so many of them in different orbits that you can always communicate between orbits.

> You can't maintain a cluster of things in orbit without actively expending propellent.

Well, or physically connecting them with tethers. (Bonus: you no longer need wireless communications between them. Malus: pretty much everything else about this design.)

One can run astonishingly accurate near-SotA ML models on even the rad hardened space rated systems with scant resources. For example after 8-bit quantization, pruning, and compression a wake-word NN can take just 14KB. An RPi tier SBC can do quite a bit of inferencing!

Hey, do you have any sources for that to read?

Sure. I saw talks by NXP and Renesas at the ARM Dev Summit last week that demonstrated the use of the Glow ML compiler [0] and CMSIS-NN Library [1] to squeeze a Wake Word NN [2] from 20KB to 14KB (with much greater reduction possible on larger ML models). Both Tensorflow Lite and Pytorch is supported in this toolchain.

[0] https://github.com/pytorch/glow

[1] https://arm-software.github.io/CMSIS_5/NN/html/index.html

[2] https://github.com/tensorflow/tensorflow/tree/master/tensorf...

Thanks, I’ll have a look. Didn’t realize ML models that small can work.

TinyML is a decent book as an intro to ML models in that size range.

Perhaps the value is more in using them as proofs of concepts? Start small, and scale up later. Those big boy space SBCs are spendy.