I'm surprised and humbled that a project of this importance doesn't have their own physical ARM machine for benchmarking. Compare that to the computer and human time saved across all the places it could be used!
They probably had their reasons, but I'm just showing what's available.
https://www.solid-run.com/product/SRLX216S00D00GE064C13CH/
https://www.solid-run.com/nxp-lx2160a-family/honeycomb-works...
I would expect benchmarks to match typical usage, if PyPy can't run in a couple of GB then choose a better language.
Don't get me wrong, i love pypy and the rpython jit.