50x isn't unreasonable if you can parallelize and/or pipeline a rigidly-defined process, cutting out as much memory and cache accesses as possible.
With the right programming model, 50x seems very, very reasonable, not even magical. A typical computer does a decently high amount of context-switching and cache-loading.