I tested 3.1alpha against 2.4 six months ago. reading large (~10^5 floats) arrays was sped up by a factor of 2, I assume due to efficiency at the xdr layer.
netcdfgroup