diff --git a/example/ck_tile/18_flatmm/flatmm_basic.cpp b/example/ck_tile/18_flatmm/flatmm_basic.cpp
index 05d0c73b7e..05d59d6e72 100644
--- a/example/ck_tile/18_flatmm/flatmm_basic.cpp
+++ b/example/ck_tile/18_flatmm/flatmm_basic.cpp
@@ -27,8 +27,8 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_con
     constexpr ck_tile::index_t N_Tile = 128;
     constexpr ck_tile::index_t K_Tile = 64;
 
-    constexpr ck_tile::index_t M_Warp = 1;
-    constexpr ck_tile::index_t N_Warp = 4;
+    constexpr ck_tile::index_t M_Warp = 2;
+    constexpr ck_tile::index_t N_Warp = 2;
     constexpr ck_tile::index_t K_Warp = 1;
 
     constexpr ck_tile::index_t M_Warp_Tile = 32;
diff --git a/include/ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1.hpp b/include/ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1.hpp
index 2d63d3b441..530eea213b 100644
--- a/include/ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1.hpp
+++ b/include/ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1.hpp
@@ -246,6 +246,8 @@ struct FlatmmPipelineAGmemBGmemCRegV1
                 });
             });
             
+            block_sync_lds();
+            
             // GEMM i
             block_flatmm(c_block_tile, a_lds_gemm_window, b_warp_tensor_2);