no convergence in NEB using damped verlet

[jruhym]

I and a colleague of mine have separately experienced the following problem.
We are running NEB simulations using the damped verlet method (IBRION = 3, I'm using SMASS = 2 and he is using the default value). Both of us are not converging. I realize that in and of itself this is not an interesting issue but our separate runs share another oddity: after each ionic step the line outputting the energy, free energy, etc. shows that the d E = F, for example
45 F= -.17728250E+04 E0= -.17727850E+04 d E =-.177283E+04 mag= 3.6094
46 F= -.17728257E+04 E0= -.17727860E+04 d E =-.177283E+04 mag= 3.6100
47 F= -.17728265E+04 E0= -.17727869E+04 d E =-.177283E+04 mag= 3.6213
48 F= -.17728262E+04 E0= -.17727859E+04 d E =-.177283E+04 mag= 3.5835
Notice that d E = F. This is to be expected after the first ionic step but certainly not after the 48th. What's going on? INCAR attached for your information.

Thanks

START = 1 # 0 don't start from WAVECAR
IMAGES = 1
SPRING = -5
ICHARG = 0 # charge from wavefunctions
#LCLIMB = .TRUE.
IBRION = 3 # = 2 # use CG method for ions
ALGO = Normal
NSW = 399
AMIX = 0.1 # Included because of electronic convergence issues
BMIX = 0.01 #3.00
#AMIN = 0.01
VOSKOWN = 1
POTIM = 0.0625
SMASS = 2.0
NPAR = 4
ISPIND = 2 # 2=spin polarized
ISPIN = 2 # 2=spin polarized
ENCUT = 400
EDIFF = 1E-5 # allowed error in total energy
EDIFFG = -0.01 # break condition for the ionic relaxation loop negtive uses forces
ISIF = 2 #relax do not include lattice parameter
ISMEAR = 0 # 0 Gaussian smearing
SIGMA = 0.05 # width of smearing, 0.05 recommended for semicond
LREAL = A
PREC = ACCURATE # Macro for precision flags
LCHARG = .FALSE. # make a CHRGCAR file

[admin]

please have a look at the total energies as written in the OUTCAR files of each image. In all vasp releases older than 4.6.32 the line F= in OSZICAR gives the sum of the total energies of all images (you can easily see it if you have a closer look at the OSZICAR file and do not just grep for the F= string: the energy of the electronic step differs SIGNIFICANTLY from the value written for F then.