Inconsistent results when subtracting two large float64 columns

[canthonissen]

test = numpy.DataFrame(numpy.random.normal(size=(80000,2)))
k = test[0] - test[1]

Works most of the time, but sometimes gives a column of zero (or near zeros) as output.
Does not seem to affect small dataframes.

[jreback]

well if you are on windows and using numexpr=2.4.4 or lower then its a dupe of: #12023

if not, pls pd.show_versions()

[canthonissen]

Yup numexpr 2.4.4.
Thanks.