diff --git a/CMakeLists.txt b/CMakeLists.txt
index f4b2ee08d943cbc852fe0ed4dcdaaf54b56349a0..5fcb0b97742d15f40d44692dc2b64a8f6bffe27a 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -155,8 +155,14 @@ if(WIN32)
     endforeach()
   endif()
 
-  # NOTE(zhouwei): msvc max/min macro conflict with std::min/max, define NOMINMAX globally
+  # msvc max/min macro conflict with std::min/max, define NOMINMAX globally
   add_definitions("-DNOMINMAX")
+
+  # 1. windows.h define 'small' cause CUDA11.6/11.7/11.8 's cub compile error,
+  # see https://github.com/microsoft/onnxruntime/issues/11227
+  # 2. WIN32_LEAN_AND_MEAN minimize the windows include files, avoid define 'small'
+  add_definitions(-DWIN32_LEAN_AND_MEAN)
+
   # windows build turn off warnings, use parallel compiling.
   foreach(
     flag_var
diff --git a/cmake/external/cub.cmake b/cmake/external/cub.cmake
index de66e8d63d069c963c98e4ecd5afb07e16dc5930..de4e04e62020213ec8a6617f485e200c0275552d 100644
--- a/cmake/external/cub.cmake
+++ b/cmake/external/cub.cmake
@@ -14,7 +14,7 @@
 
 include(ExternalProject)
 
-# Note(zhouwei): extern_cub  has code __FILE_, If the path of extern_cub is changed,
+# extern_cub  has code __FILE_, If the path of extern_cub is changed,
 # it will effect about 30+ cu files sccache hit and slow compile speed  on windows.
 # Therefore, a fixed CUB_PATH will be input to increase the sccache hit rate.
 set(CUB_PATH
@@ -25,7 +25,7 @@ set(CUB_PREFIX_DIR ${CUB_PATH})
 set(CUB_REPOSITORY ${GIT_URL}/NVlabs/cub.git)
 
 if(${CMAKE_CUDA_COMPILER_VERSION} GREATER_EQUAL 11.6)
-  # cuda_11.6.2_511.65‘s own cub is 1.15.0, which will cause compiling error in windows.
+  # cuda_11.6/11.7/11.8‘s own cub is 1.15.0, which will cause compiling error in windows.
   set(CUB_TAG 1.16.0)
   # cub 1.16.0 is not compitable with current thrust version
   add_definitions(-DTHRUST_IGNORE_CUB_VERSION_CHECK)
diff --git a/cmake/third_party.cmake b/cmake/third_party.cmake
index 02c5bc3602f737f3dea9777971fb49b4e3682d87..f4284c8421428a4440e94b1ca84eda6f83b81845 100755
--- a/cmake/third_party.cmake
+++ b/cmake/third_party.cmake
@@ -317,8 +317,7 @@ if(WITH_ONNXRUNTIME)
 endif()
 
 if(WITH_GPU)
-  if(${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0 OR ${CMAKE_CUDA_COMPILER_VERSION}
-                                                 GREATER_EQUAL 11.6)
+  if(${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0)
     include(external/cub) # download cub
     list(APPEND third_party_deps extern_cub)
   endif()