In my experience 'small-batch distributed jobs' (~5 machines) are often candidates for memory optimization. Especially if written in high-level languages without easy access to struct packing.
It boggles me that a machine that could hold six hundred uncompressed 1080p video frames is my most memory constrained computer.
I'm on board with Niklaus Wirth: https://cr.yp.to/bib/1995/wirth.pdf
So the results will probably be pretty similar on ARM as well.
Memory consumption can be a primary issue for small embedded targets. Upgrading the hardware is often not an option or is actually more costly than development time, because development is a one-time cost, while adding RAM or flash is an extra production cost (so a per-unit cost).
Moreover, embedded software tends to be more stable feature-wise than desktop or server software: once it's done, you can expect that it won't be modified for a few years. From my experience, the smaller the system the more "carved in stone" it is.
Finally, it should be noticed than on "normal" systems, memory size is still important because of the many caches they feature.
http://www.7-cpu.com/cpu/Skylake.html
L1 cache: 32 KiB, 4 clock cycles (=1 nanosecond) latency
L3 cache: (Up to) 8 MiB… with a latency (42ns) equivalent of 2½ branch mispredictions. You know, the things that Everyone Knows™ they should avoid with compiler hints like __builtin_expect.
RAM: As much as you want… at 246 cycles (61.5ns) latency.
That said, understanding how to design in tight memory constraints is useful. While typically only seriously practiced by embedded systems developers, having habits that minimize memory use can have a large impact at scale. The paper gives some good reasons on what the benefits of those habits are, but I also think that, as a percentage of the total, programmers living in constrained memory spaces are a specialization not the mainstream any more.