These days I'm pushing a few changes to slightly improve performance of the conversions, and others will come in the next days ;-)

In addition, there is a one-line change that can boost performance: enabling multi-threading

diff --git a/src/SkyCoords.jl b/src/SkyCoords.jl
index 5aa77d2..a092321 100644
--- a/src/SkyCoords.jl
+++ b/src/SkyCoords.jl
@@ -258,7 +258,7 @@ function convert{T<:AbstractSkyCoords, n, S<:AbstractSkyCoords}(
     ::Type{Array{T,n}}, c::Array{S, n})
     m = Matrix33{_eltype(S)}(rotmat(T, S))
     result = similar(c, T)
-    for i in 1:length(c)
+    Threads.@threads for i in 1:length(c)
         r = m * coords2cart(c[i])
         lon, lat = cart2coords(r)
         result[i] = T(lon, lat)

Running the benchmark script at https://github.com/kbarbary/SkyCoords.jl/pull/9#issuecomment-336395556 I get this new plot by using 4 threads:

With 1 million coordinates SkyCoords.jl is still one order of magnitude faster than Astropy :-) The change was applied on top of current master, so there is still room for some minor improvement.

The problem of multi-threading in Julia is that it's still experimental. However, when running Julia with just one thread there is zero overhead and no issues. Thus, the change I propose should be safe enough (doesn't break single-thread systems) and we would only need to warn people that multi-threading may not work.