- Time/margin tradeoffs: if the attacker can only meaningfully attack in the next 24 hours, do you need margin to protect against all of AWS crunching for a million years? Or would you like the computation done in a tenth the time?
- Architectural properties, such as the susceptibility to differential power analysis: maybe you will risk that more via a LUT in exchange for lower power.
- As you increase the number of rounds, algorithms tend to get stronger. But some algorithms have E.g. good avalanche properties for many fewer rounds.
The tiniest devices are things like RFID chips.
For example it could have low RAM usage, but become exponentially slower with message size, so it's fine for encrypting short messages only.
Or it could have a constant overhead somewhere that's fine for an IoT device that sends a report once a day, but not if you wanted to make thousands of connections per second.
Stuff like that, is my naive understanding.