diff --git a/cmake/modules/CUDA.cmake b/cmake/modules/CUDA.cmake
index 1284f85bec..1ebf4e2234 100644
--- a/cmake/modules/CUDA.cmake
+++ b/cmake/modules/CUDA.cmake
@@ -109,6 +109,30 @@ if(USE_CUDA)
   # Add CUDA builtins to RelaxVM
   tvm_file_glob(GLOB RELAX_VM_CUDA_BUILTIN_SRC_CC src/runtime/relax_vm/cuda/*.cc)
   list(APPEND RUNTIME_SRCS ${RELAX_VM_CUDA_BUILTIN_SRC_CC})
+
+  if(USE_CUTLASS)
+    if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+      set(CMAKE_CUDA_COMPILER ${CUDA_TOOLKIT_ROOT_DIR}/bin/nvcc)
+      execute_process(
+        COMMAND ${CMAKE_CUDA_COMPILER} --version
+        OUTPUT_VARIABLE NVCC_VERSION_OUTPUT
+      )
+      if(NVCC_VERSION_OUTPUT MATCHES "release ([^,]+),")
+        set(CUDA_VERSION "${CMAKE_MATCH_1}")
+      endif(NVCC_VERSION_OUTPUT MATCHES "release ([^,]+),")
+      message(STATUS "CUDA_VERSION=${CUDA_VERSION}")
+
+      if(CUDA_VERSION VERSION_GREATER_EQUAL 11.8)
+        set(CMAKE_CUDA_ARCHITECTURES "75;80;86;89;90")
+      else(CUDA_VERSION VERSION_GREATER_EQUAL 11.8)
+        set(CMAKE_CUDA_ARCHITECTURES "75;80;86")
+      endif(CUDA_VERSION VERSION_GREATER_EQUAL 11.8)
+      message(STATUS "Set CMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}")
+    else(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+      message(STATUS "Found CMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}")
+    endif(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+  endif(USE_CUTLASS)
+
 else(USE_CUDA)
   list(APPEND COMPILER_SRCS src/target/opt/build_cuda_off.cc)
 endif(USE_CUDA)