I am trying to use numpy‘s nbytes attribute to examine the memory usage of arrays with different dtype. I noticed the following:

>>> np.zeros(1024, dtype='int64').nbytes / 1024    # in kB
8.0
>>> np.zeros(1024, dtype='int8').nbytes / 1024
1.0
>>> np.zeros(1024, dtype='bool').nbytes / 1024
1.0  # ?

The results of the first two lines make sense to me, since the memory usage reported (8 kB and 1 kB, resp.) is simply the sum of the memory usage of each element (8 and 1 bytes, resp.) multiplied by the number of elements (1024 in both cases). However, in the last case, I would assume that an array of bool type only requires one bit of memory per element, however in seems that it requires 8 bits per element, same as int8.

My questions:

• Am I interpreting correctly the output of nbytes? If so, how should we understand the fact that a bool type does not use one bit per element?
• With this in mind, is there a reason to use a bool type array over an int8?

This question came up as I am dealing with arrays of binary (0 and 1) values, and was trying to find the space efficient solution (in terms of memory, since I assume there are not differences in computation times?), which I thought would be using a bool type.