I don't find that compelling. Compare:
Original + restart
Loop time of 506.873 on 1 procs for 10000000 steps with 101 atoms
Loop time of 1676.44 on 1 procs for 10000000 steps with 101 atoms
Original with 2*NT0
Loop time of 968.009 on 1 procs for 20000000 steps with 101 atoms
While there is certainly some batch to batch variation and effect of the i/o, the total time is still out of line.