When properly configured, most of these libraries use NVidia's CuDNN package under the hood. The only thing you're really measuring here is overhead, not the actual computation.