diff --git a/Grid/lattice/Lattice_reduction.h b/Grid/lattice/Lattice_reduction.h
index c3478ab4..0ddac437 100644
--- a/Grid/lattice/Lattice_reduction.h
+++ b/Grid/lattice/Lattice_reduction.h
@@ -168,6 +168,22 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
   return ssum;
 }
 
+template<class vobj>
+inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+{
+#if defined(GRID_CUDA)||defined(GRID_HIP)
+  autoView( arg_v, arg, AcceleratorRead);
+  Integer osites = arg.Grid()->oSites();
+  auto ssum= sum_gpu_large(&arg_v[0],osites);
+#else
+  autoView(arg_v, arg, CpuRead);
+  Integer osites = arg.Grid()->oSites();
+  auto ssum= sum_cpu(&arg_v[0],osites);
+#endif
+  arg.Grid()->GlobalSum(ssum);
+  return ssum;
+}
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Deterministic Reduction operations
 ////////////////////////////////////////////////////////////////////////////////////////////////////
diff --git a/Grid/lattice/Lattice_reduction_gpu.h b/Grid/lattice/Lattice_reduction_gpu.h
index c685a2c0..c3422af3 100644
--- a/Grid/lattice/Lattice_reduction_gpu.h
+++ b/Grid/lattice/Lattice_reduction_gpu.h
@@ -288,6 +288,14 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
   return result;
 }
 
+template <class vobj>
+inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
+{
+  typedef typename vobj::scalar_object sobj;
+  sobj result;
+  result = sumD_gpu_large(lat,osites);
+  return result;
+}
 
 
 NAMESPACE_END(Grid);