By default. 3dUndump creates a dataset with "short" integers (16 bit precision). So an input of 12.3 would be turned into 12, losing the .3 forever. To avoid this issue, use the '-datum float' option on the command line. For more details, read the -help output in detail.
This choice of 'datum' (the storage format for the values at each voxel) affect other programs, especially 3dcalc.