diff -Nr -u5 pmemd_orig/src/remd_exchg.F90 pmemd_dump/src/remd_exchg.F90
--- pmemd_orig/src/remd_exchg.F90	2012-12-15 19:57:56.000000000 +0100
+++ pmemd_dump/src/remd_exchg.F90	2012-12-16 21:04:24.000000000 +0100
@@ -436,10 +436,15 @@
   type(pme_pot_ene_rec) :: my_new_pme_ene ! pme energy record for THEIR coords
   type(ene_temp)   :: my_ene_temp
   type(ene_temp)   :: neighbor_ene_temp
   type(ene_temp)   :: ene_temp_buffer(numgroups) ! to collect/print REMD data
 
+  logical, save :: force_not_dumped = .true.
+  logical, save :: force_not_reduced = .true.
+  double precision, dimension(:,:), allocatable :: f_buff
+  character(80) dump_name
+
 
   ! Get coordinates and forces from GPU
 #ifdef CUDA
   call gpu_download_crd(crd)
   call gpu_download_frc(frc)
@@ -524,10 +529,38 @@
                    my_atm_lst, .true., .true., .false., my_new_pme_ene, &
                    virial, ekcmt, pme_err_est)
     
     my_ene_temp%energy_2 = my_new_pme_ene%total
 
+    if (force_not_reduced) then
+       allocate(f_buff(3, atm_cnt))
+       call MPI_Reduce(frc_temp, f_buff, atm_cnt*3, MPI_DOUBLE_PRECISION, &
+                       MPI_SUM, 0, pmemd_comm, err_code_mpi)
+       force_not_reduced = .false.
+    end if
+    if (force_not_dumped) then
+       if (master) then
+          write(6, '(/"dumping neighbour forces"/)')
+          !write(dump_name, '("array_dump."i3.3)') master_rank + 1
+          !open(95, file = dump_name, status = 'old', position = 'append', action = 'write')
+          write(95, '("neighbour coordinates")')
+          write(95, *) crd_temp
+          write(95, '("end")')
+          write(95, '("neighbour forces")')
+          write(95, *) frc_temp
+          write(95, '("end")')
+          write(95, '("neighbour forces after reduction")')
+          write(95, *) f_buff
+          write(95, '("end")')
+          close(95)
+          deallocate(f_buff)
+       else
+          deallocate(f_buff)
+       end if
+       force_not_dumped = .false.
+    end if
+
   end if
 
 
 #ifdef CUDA
   ! Update the results from the GPU
diff -Nr -u5 pmemd_orig/src/runmd.F90 pmemd_dump/src/runmd.F90
--- pmemd_orig/src/runmd.F90	2012-12-15 19:57:56.000000000 +0100
+++ pmemd_dump/src/runmd.F90	2012-12-16 21:04:03.000000000 +0100
@@ -101,10 +101,15 @@
 #ifdef MPI
   double precision              :: remd_ptot
   double precision              :: new_list_cnt
   double precision, save        :: reduce_buf_in(7)
   double precision, save        :: reduce_buf_out(7)
+
+  logical, save :: force_not_dumped = .true.
+  logical, save :: force_not_reduced = .true.
+  double precision, dimension(:,:), allocatable :: f_buff
+  character(80) dump_name
 #endif
   double precision      :: fac(3)
   double precision      :: rmu(3)
   double precision      :: factt
   double precision      :: pconv
@@ -604,10 +609,38 @@
                        my_atm_lst, new_list, need_pot_enes, need_virials, &
                        pme_pot_ene, virial, ekcmt, pme_err_est)
 
         remd_ptot = pme_pot_ene%total
 
+        if (force_not_reduced) then
+           allocate(f_buff(3, atm_cnt))
+           call MPI_Reduce(frc, f_buff, atm_cnt*3, MPI_DOUBLE_PRECISION, &
+                           MPI_SUM, 0, pmemd_comm, err_code_mpi)
+           force_not_reduced = .false.
+        end if
+        if (force_not_dumped) then
+           if (master) then
+              write(6, '(/"dumping own forces"/)')
+              write(dump_name, '("array_dump."i3.3)') master_rank + 1
+              open(95, file = dump_name, status = 'replace', action = 'write')
+              write(95, '("own coordinates")')
+              write(95, *) crd
+              write(95, '("end")')
+              write(95, '("own forces")')
+              write(95, *) frc
+              write(95, '("end")')
+              write(95, '("own forces after reduction")')
+              write(95, *) f_buff
+              write(95, '("end")')
+              !close(95)
+              deallocate(f_buff)
+           else
+              deallocate(f_buff)
+           end if
+           force_not_dumped = .false.
+        end if
+
       else if (using_gb_potential) then
         call gb_force(atm_cnt, crd, frc, gb_pot_ene, irespa, need_pot_enes)
 
         remd_ptot = gb_pot_ene%total