diff --git a/app/celer-g4/GlobalSetup.cc b/app/celer-g4/GlobalSetup.cc
index 36f4648c28..f616713efb 100644
--- a/app/celer-g4/GlobalSetup.cc
+++ b/app/celer-g4/GlobalSetup.cc
@@ -167,6 +167,7 @@ void GlobalSetup::ReadInput(std::string const& filename)
         options_->cuda_heap_size = input_.cuda_heap_size;
         options_->sync = input_.sync;
         options_->default_stream = input_.default_stream;
+        options_->auto_flush = input_.auto_flush;
     }
     else if (ends_with(filename, ".mac"))
     {
diff --git a/app/celer-g4/RunInput.cc b/app/celer-g4/RunInput.cc
index 157592574b..67b1e74e02 100644
--- a/app/celer-g4/RunInput.cc
+++ b/app/celer-g4/RunInput.cc
@@ -51,7 +51,8 @@ RunInput::operator bool() const
            && physics_list < PhysicsListSelection::size_
            && (field == no_field() || field_options)
            && ((num_track_slots > 0 && max_steps > 0
-                && initializer_capacity > 0 && secondary_stack_factor > 0)
+                && initializer_capacity > 0 && secondary_stack_factor > 0
+                && auto_flush > 0 && auto_flush <= initializer_capacity)
                || SharedParams::CeleritasDisabled())
            && (step_diagnostic_bins > 0 || !step_diagnostic);
 }
diff --git a/app/celer-g4/RunInput.hh b/app/celer-g4/RunInput.hh
index 99f48bc1c6..d8c17adee1 100644
--- a/app/celer-g4/RunInput.hh
+++ b/app/celer-g4/RunInput.hh
@@ -67,6 +67,7 @@ struct RunInput
     size_type max_steps{unspecified};
     size_type initializer_capacity{};
     real_type secondary_stack_factor{};
+    size_type auto_flush{};  //!< Defaults to num_track_slots
     bool sync{false};
     bool default_stream{false};  //!< Launch all kernels on the default stream
 
diff --git a/app/celer-g4/RunInputIO.json.cc b/app/celer-g4/RunInputIO.json.cc
index b14bcd7f32..10130f2301 100644
--- a/app/celer-g4/RunInputIO.json.cc
+++ b/app/celer-g4/RunInputIO.json.cc
@@ -72,6 +72,14 @@ void from_json(nlohmann::json const& j, RunInput& v)
     RI_LOAD_OPTION(secondary_stack_factor);
     RI_LOAD_OPTION(sync);
     RI_LOAD_OPTION(default_stream);
+    if (auto iter = j.find("auto_flush"); iter != j.end())
+    {
+        iter->get_to(v.auto_flush);
+    }
+    else
+    {
+        v.auto_flush = v.num_track_slots;
+    }
 
     RI_LOAD_OPTION(physics_list);
     RI_LOAD_OPTION(physics_options);
@@ -162,6 +170,7 @@ void to_json(nlohmann::json& j, RunInput const& v)
     RI_SAVE_OPTION(cuda_heap_size);
     RI_SAVE(sync);
     RI_SAVE(default_stream);
+    RI_SAVE(auto_flush);
 
     RI_SAVE(physics_list);
     if (v.physics_list == PhysicsListSelection::geant_physics_list)
diff --git a/src/accel/LocalTransporter.cc b/src/accel/LocalTransporter.cc
index 270b345929..52743ed2ba 100644
--- a/src/accel/LocalTransporter.cc
+++ b/src/accel/LocalTransporter.cc
@@ -50,7 +50,8 @@ namespace celeritas
  */
 LocalTransporter::LocalTransporter(SetupOptions const& options,
                                    SharedParams const& params)
-    : auto_flush_(options.max_num_tracks)
+    : auto_flush_(options.auto_flush ? options.auto_flush
+                                     : options.max_num_tracks)
     , max_steps_(options.max_steps)
     , dump_primaries_{params.offload_writer()}
     , hit_manager_{params.hit_manager()}
diff --git a/src/accel/SetupOptions.hh b/src/accel/SetupOptions.hh
index 44df17d859..d3d24aa8b2 100644
--- a/src/accel/SetupOptions.hh
+++ b/src/accel/SetupOptions.hh
@@ -126,6 +126,8 @@ struct SetupOptions
     size_type initializer_capacity{};
     //! At least the average number of secondaries per track slot
     real_type secondary_stack_factor{3.0};
+    //! Number of tracks to buffer before offloading (if unset: max num tracks)
+    size_type auto_flush{};
     //!@}
 
     //! Set the number of streams (defaults to run manager # threads)
diff --git a/src/accel/SetupOptionsMessenger.cc b/src/accel/SetupOptionsMessenger.cc
index 3be3b41ce6..7c51cf226f 100644
--- a/src/accel/SetupOptionsMessenger.cc
+++ b/src/accel/SetupOptionsMessenger.cc
@@ -190,6 +190,9 @@ SetupOptionsMessenger::SetupOptionsMessenger(SetupOptions* options)
     add_cmd(&options->secondary_stack_factor,
             "secondaryStackFactor",
             "At least the average number of secondaries per track slot");
+    add_cmd(&options->auto_flush,
+            "autoFlush",
+            "Number of tracks to buffer before offloading");
 
     directories_.emplace_back(new CelerDirectory(
         "/celer/detector/", "Celeritas sensitive detector setup options"));
diff --git a/src/accel/SetupOptionsMessenger.hh b/src/accel/SetupOptionsMessenger.hh
index 60344b306e..e4657f65f9 100644
--- a/src/accel/SetupOptionsMessenger.hh
+++ b/src/accel/SetupOptionsMessenger.hh
@@ -33,6 +33,7 @@ struct SetupOptions;
   maxNumSteps          | Limit on number of step iterations before aborting
   maxInitializers      | Maximum number of track initializers
   secondaryStackFactor | At least the average number of secondaries per track
+  autoFlush            | Number of tracks to buffer before offloading
 
  * The following option is exposed in the \c /celer/detector/ command
  * "directory":