Our iterator avoids buffers whenever possible, this is the case as usually accessing data with strides is faster than copying to a buffer for a contigous loop. This is also the case for our vectorized basic math functions (+-*/).

Since we merged vectorized exp and log we have vectorized ufuncs that are heavily cpu bound instead of memory bound.
For these functions it can be very worthwhile to use buffered iterators as only then the vectorized code can be used.
Another advantage is that it we guarantee the same results regardless of strides of the input data. (A property that in current master might be lost but could also be restored without a buffered iterator by using the vector code on scalar data though that might be a bit slower)

import numpy as np
d = np.random.rand(1000, 50).astype(np.float32)
print("2d strided -> buffered iterator")
%timeit np.exp(d[::2])
d = np.random.rand(1000* 50).astype(np.float32)
print("1d strided unbuffered iterator")
%timeit np.exp(d[::2])


2d strided -> buffered iterator
10000 loops, best of 3: 77.3 µs per loop
1d strided unbuffered iterator
1000 loops, best of 3: 1.02 ms per loop

Btw vectorizing the functions for strided data adds complexity and would likely be slower than our buffer approach. I would not recommend pursuing it.