diff --git a/doc/src/Speed_kokkos.rst b/doc/src/Speed_kokkos.rst
index 569a24f1c2..dd417d7c79 100644
--- a/doc/src/Speed_kokkos.rst
+++ b/doc/src/Speed_kokkos.rst
@@ -285,6 +285,16 @@ one or more nodes, each with two GPUs:
    settings. Experimenting with its options can provide a speed-up for
    specific calculations. For example:
 
+.. note::
+
+   The default binsize for :doc:`atom sorting <atom_modify>` on GPUs
+   is equal to the default CPU neighbor binsize (i.e. 2x smaller than the
+   default neighbor binsize on GPUs). When running simple pair-wise
+   potentials like Lennard Jones on GPUs, using a 2x larger binsize for
+   atom sorting (equal to the default binsize for building the neighbor
+   list on GPUs) and a more frequent sorting than default (e.g. sorting
+   every 100 time steps instead of 1000) may improve performance.
+
 .. code-block:: bash
 
    mpirun -np 2 lmp_kokkos_cuda_openmpi -k on g 2 -sf kk -pk kokkos newton on neigh half binsize 2.8 -in in.lj      # Newton on, half neighbor list, set binsize = neighbor ghost cutoff
diff --git a/doc/src/atom_modify.rst b/doc/src/atom_modify.rst
index 9049a24fde..1e5a3d49ff 100644
--- a/doc/src/atom_modify.rst
+++ b/doc/src/atom_modify.rst
@@ -153,6 +153,13 @@ cache locality will be undermined.
    order of atoms in a :doc:`dump <dump>` file will also typically change
    if sorting is enabled.
 
+.. note::
+
+   When running simple pair-wise potentials like Lennard Jones on GPUs
+   with the KOKKOS package, using a larger binsize (e.g. 2x larger than
+   default) and a more frequent reordering than default (e.g. every 100
+   time steps) may improve performance.
+
 Restrictions
 """"""""""""