Across millions of tries, same thing: Windows exp2 is off by at least 1 ulp over a third of the time, and by over 2 ulp about 3 times per million. Still haven't seen pow(2, x) off by as much as 0.52 ulp.

From its behavior, it appears Windows implements exp2(x) like so:

    i = floor(x)
    x -= i # now 0 <= x < 1
    return ldexp(exp2(x), i)

So it's apparently using some sub-state-of-the-art approximation to 2**x over the domain [0, 1].

But a consequence is that it gets it exactly right whenever x is an integer, so it's unlikely anyone will notice it's sloppy ;-)

I expect we should just live with it.