Tbh, i didn't analyse the two constructs too see which one is better performing. Thanks for doing that 
Btw i would say that i'm surprised too how can the internal table approach be faster, some kernel magic dust 
But for sure the memory consumption is more with the internal table, hopefully.
BR,
Suhas
PS - Can you post the code snippet you used for your testing? I can run a trace in my system & validate
