diff --git a/cmake/create_test.cmake b/cmake/create_test.cmake
index 9d22406f9ac..dcc452b293b 100644
--- a/cmake/create_test.cmake
+++ b/cmake/create_test.cmake
@@ -40,6 +40,7 @@ function(ginkgo_create_dpcpp_test test_name)
     add_executable(${test_target_name} ${test_name}.dp.cpp)
     target_compile_features(${test_target_name} PUBLIC cxx_std_17)
     target_compile_options(${test_target_name} PRIVATE "${GINKGO_DPCPP_FLAGS}")
+    target_compile_options(${test_target_name} PRIVATE "${GINKGO_COMPILER_FLAGS}")
     target_link_options(${test_target_name} PRIVATE -fsycl-device-code-split=per_kernel)
     ginkgo_set_test_target_properties(${test_name} ${test_target_name})
     # Note: MKL_ENV is empty on linux. Maybe need to apply MKL_ENV to all test.
diff --git a/dpcpp/CMakeLists.txt b/dpcpp/CMakeLists.txt
index fee9ec36396..dc37a35ebd3 100644
--- a/dpcpp/CMakeLists.txt
+++ b/dpcpp/CMakeLists.txt
@@ -66,8 +66,11 @@ target_sources(ginkgo_dpcpp
 ginkgo_compile_features(ginkgo_dpcpp)
 target_compile_definitions(ginkgo_dpcpp PRIVATE GKO_COMPILING_DPCPP)
 
+# -sycl-std=2017 is temporary workaround after 2021.4 to propagate subgroup setting correctly
+set(GINKGO_DPCPP_FLAGS "-sycl-std=2017")
 set(GINKGO_DPCPP_FLAGS ${GINKGO_DPCPP_FLAGS} PARENT_SCOPE)
 target_compile_options(ginkgo_dpcpp PRIVATE "${GINKGO_DPCPP_FLAGS}")
+target_compile_options(ginkgo_dpcpp PRIVATE "${GINKGO_COMPILER_FLAGS}")
 # Note: add MKL as PRIVATE not PUBLIC (MKL example shows) to avoid propagating
 # find_package(MKL) everywhere when linking ginkgo (see the MKL example
 # https://software.intel.com/content/www/us/en/develop/documentation/onemkl-windows-developer-guide/top/getting-started/cmake-config-for-onemkl.html)
diff --git a/dpcpp/matrix/dense_kernels.dp.cpp b/dpcpp/matrix/dense_kernels.dp.cpp
index 0c89530d1d2..9a86ab9cd15 100644
--- a/dpcpp/matrix/dense_kernels.dp.cpp
+++ b/dpcpp/matrix/dense_kernels.dp.cpp
@@ -69,14 +69,14 @@ namespace dpcpp {
 namespace dense {
 
 
+// Disable the 64 subgroup. CPU supports 64 now, but conj_transpose will
+// lead CL_OUT_OF_RESOURCES. TODO: investigate this issue.
 using KCFG_1D = ConfigSet<11, 7>;
 constexpr auto kcfg_1d_list =
-    syn::value_list<std::uint32_t, KCFG_1D::encode(512, 64),
-                    KCFG_1D::encode(512, 32), KCFG_1D::encode(512, 16),
-                    KCFG_1D::encode(256, 32), KCFG_1D::encode(256, 16),
-                    KCFG_1D::encode(256, 8)>();
-constexpr auto subgroup_list =
-    syn::value_list<std::uint32_t, 64, 32, 16, 8, 4>();
+    syn::value_list<std::uint32_t, KCFG_1D::encode(512, 32),
+                    KCFG_1D::encode(512, 16), KCFG_1D::encode(256, 32),
+                    KCFG_1D::encode(256, 16), KCFG_1D::encode(256, 8)>();
+constexpr auto subgroup_list = syn::value_list<std::uint32_t, 32, 16, 8, 4>();
 constexpr auto kcfg_1d_array = syn::as_array(kcfg_1d_list);
 constexpr int default_block_size = 256;