What are the differences between the IBM machines?

Besides number of qubits, the devices can have other differences as well. The architecture of the device can be different, meaning that each device could have different connectivity maps. This would affect the mapping of valid multiqubit gates.

They also can have different error rates at any given time. Calibrations are run on each device daily. These error rates can change after each calibration. If one device has poor error rates compared to another device, it's accuracy will most likely be less than the device with the better error rates.

Those are some high level differences at least.

The inherent uncertainty of QM is managed by making sure you're taking enough shots relative to the error bounds and probability of success for a given algorithm. If you don't have exact equations, and haven't done so already, just watching the extent to which your results convergence, or fail to converge, as you increase your shots, is a good first step.

Regarding differences between hardware, some information is aggregated here. But you can also map your logical qubits to the physical qubits of specific hardware, at specific optimization levels directly in Qiskit. This Q&A is a good example with corresponding code.

EDIT: Some of the links in the linked Q&A are apparently dead. This pulse tutorial has a lot of good information on getting backend information in Qiskit. You may also find this tutorial on transpiler passes and optimization levels relevant.

First of all the name of backends (devices) has nothing to do with their location! They are all located in US.

Back to your question, as others already mentioned the difference is in the architecture (topology) and the error rates.

If you click the name of any backend (device) in your IBM_Q Experience dashboard, you'll see detailed information about the error rate for each architecture.

For example ibmq_london and ibmq_essex have the same topology but quit_1 in ibmq_london has lower error rate (noise) compared to the same qubit in ibmq_essex. That's the reason applying the same gate to same qubits on different backends, will result in different output.