This paper addresses the question of the weakest failure detector for solving the Non-Blocking Atomic Commit problem (NBAC) in a message passing system where processes can fail by crashing. We define a failure detector, denoted by X, which we show to be sufficient to solve NBAC with a majority of correct processes. Then we give an algorithm which, no matter how many processes may crash, uses any failure detector that solves NBAC to emulate X, i.e., we prove that \X~is also necessary for solving NBAC. The result was obtained concurrently and independently by Hadzilacos and Toueg.