If you want to learn clustering I would start with a couple of cheap Orange Pi One or PC. Or wait for the first reviews of Banana Pi M2+ (but there's not much to test since it's more or less a clone of Orange Pi Plus regarding hardware settings and pin mappings, so you get Gbit Ethernet and eMMC as with the M3 but since H3 is pretty good supported by 3rd party distros you don't have to use the somewhat... horrible... OS images provided here)
Regarding M3, heat and peripherals you should read again. If you improve heat dissipation (and you'll need a strong fan + heatsinks) one then M3 will show superiour performance and you also have to take care that you can provide ~10W per board.
If your cluster workloads are CUDA capable then relying on slow ARM clusters is just a bad joke (if you really need performance and do not only want to learn building clusters). A cheap PC using one single rather cheap NVIDIA GTX 770 card comes with 1536 CUDA capable cores running at +1GHz. You will need a few hundred Bananas, network switches like crazy, liquid cooling and a bunch of fat PSUs... and won't be faster.
Clustering is nice for a few use cases, but if your problems are 'GPGPU capable' then it's a bad idea. Or the other way around: Using a server mainboard where you use a couple fat NVIDIA GPUs is always the better idea.