We should add some set of standard Numpy benchmarks to pyperformance. Numpy itself includes official microbenchmarks, but as a starting point, it's probably more useful to use the externally-maintained macrobenchmarks at https://github.com/spcl/npbench. While the primary purpose of these is to compare alternative approaches to numeric computation (numpy, numba, pythran etc.), as a starting point, just benchmarking in numpy should be sufficient.

I wouldn't expect most of the work thus far to have a significant impact on these benchmarks, as they are largely exercising code in C extensions. However, I think these represent important workloads to confirm that we aren't inadvertently introducing regressions. And having these in place may be useful if/when we focus on the performance of Python<->C data structure marshalling etc.