I've looked into similair cases, but I've yet to find one where the cluster can share both WUs and memory (e.g. work on one WCG WU over 12 cores by using the RAM of the entire 40-board cluster)
It is amazing to see how fast a question as "What would you do with a 120-Raspberry Pi Cluster?"
can be translated into "What would you do with a 15-Banana Pi M3 Cluster?" and getting more performance for less wattage.
I hadn't yet, but they are small in both size (NanoPi-2 and NanoPi-2-Fire are hardly bigger than a Raspberry Pi Zero)
and price (The quad-core Allwinner H3 NanoPi-M1 can be bought for a price as low as 13 US$, the octo-core Nanopi-M3 is,
together with the NanoPi-2 and NanoPi-2-Fire, the most expensive at only 32 US$)
The models M1, M2 and M3 all slightly differ in size (64 mm x 50-56-60 mm, but are generally smaller than the Raspberry Pi A+ (65 mm x 56.5 mm).