diff --git a/include/dlaf/eigensolver/bt_band_to_tridiag/impl.h b/include/dlaf/eigensolver/bt_band_to_tridiag/impl.h
index 49de0188c0..50a178ff3b 100644
--- a/include/dlaf/eigensolver/bt_band_to_tridiag/impl.h
+++ b/include/dlaf/eigensolver/bt_band_to_tridiag/impl.h
@@ -631,7 +631,7 @@ void BackTransformationT2B<B, D, T>::call(const SizeType band_size,
   const LocalTileSize tiles_per_block_hh(mat_hh.blockSize().rows() / b, mat_hh.blockSize().cols() / b);
   Matrix<const T, Device::CPU> mat_hh_rt = mat_hh.retiledSubPipelineConst(tiles_per_block_hh);
   const LocalTileSize tiles_per_block_e(mat_e.blockSize().rows() / b, 1);
-  matrix::internal::MatrixRef<T, D> mat_e_rt = mat_e.retiledSubPipeline(tiles_per_block_e);
+  Matrix<T, D> mat_e_rt = mat_e.retiledSubPipeline(tiles_per_block_e);
 
   const auto& dist_hh_rt = mat_hh_rt.distribution();
   const auto& dist_e_rt = mat_e_rt.distribution();
@@ -760,7 +760,7 @@ void BackTransformationT2B<B, D, T>::call(comm::CommunicatorGrid& grid, const Si
   const SizeType group_size = getTuneParameters().bt_band_to_tridiag_hh_apply_group_size;
 
   const LocalTileSize tiles_per_block(mat_e.blockSize().rows() / b, 1);
-  matrix::internal::MatrixRef<T, D> mat_e_rt = mat_e.retiledSubPipeline(tiles_per_block);
+  Matrix<T, D> mat_e_rt = mat_e.retiledSubPipeline(tiles_per_block);
 
   const auto& dist_hh = mat_hh.distribution();
   const auto& dist_e_rt = mat_e_rt.distribution();
diff --git a/include/dlaf/matrix/matrix_ref.h b/include/dlaf/matrix/matrix_ref.h
index cbcbb50328..1302c78b4b 100644
--- a/include/dlaf/matrix/matrix_ref.h
+++ b/include/dlaf/matrix/matrix_ref.h
@@ -55,10 +55,6 @@ class MatrixRef<const T, D> : public internal::MatrixBase {
       : internal::MatrixBase(Distribution(mat.distribution(), spec)), mat_const_(mat),
         origin_(spec.origin) {}
 
-  MatrixRef(Matrix<const T, D>&& mat, const SubMatrixSpec& spec)
-      : internal::MatrixBase(Distribution(mat.distribution(), spec)), mat_const_(mat),
-        origin_(spec.origin) {}
-
   /// Create a matrix reference of @p mat.
   ///
   /// @param[in] mat is the input matrix.
@@ -79,9 +75,9 @@ class MatrixRef<const T, D> : public internal::MatrixBase {
   ///
   /// @pre blockSize() is divisible by @p tiles_per_block
   /// @pre blockSize() == tile_size()
-  MatrixRef<const T, D> retiledSubPipelineConst(const LocalTileSize& tiles_per_block) {
-    SubMatrixSpec spec = {this->distribution().origin(), this->distribution().size()};
-    return MatrixRef<const T, D>(std::move(mat_const_.retiledSubPipelineConst(tiles_per_block)), spec);
+  Matrix<const T, D> retiledSubPipelineConst(const LocalTileSize& tiles_per_block) {
+    // SubMatrixSpec spec = {this->distribution().origin(), this->distribution().size()};
+    return mat_const_.retiledSubPipelineConst(tiles_per_block);
   }
 
   /// Returns a read-only sender of the Tile with local index @p index.
@@ -173,9 +169,9 @@ class MatrixRef : public MatrixRef<const T, D> {
   ///
   /// @pre blockSize() is divisible by @p tiles_per_block
   /// @pre blockSize() == tile_size()
-  MatrixRef<T, D> retiledSubPipeline(const LocalTileSize& tiles_per_block) noexcept {
-    SubMatrixSpec spec = {this->distribution().origin(), this->distribution().size()};
-    return MatrixRef<T, D>(std::move(mat_.retiledSubPipeline(tiles_per_block)), spec);
+  Matrix<T, D> retiledSubPipeline(const LocalTileSize& tiles_per_block) noexcept {
+    // SubMatrixSpec spec = {this->distribution().origin(), this->distribution().size()};
+    return mat_.retiledSubPipeline(tiles_per_block);
   }
 
   /// Returns a sender of the Tile with local index @p index.