diff --git a/cmake/external/xpu.cmake b/cmake/external/xpu.cmake
index 138f06c4ae8fc531096cb45aa3c9d71b177a6194..65c02e1e520a8e0a0611eaf475e8a93b99a8aa03 100644
--- a/cmake/external/xpu.cmake
+++ b/cmake/external/xpu.cmake
@@ -8,7 +8,7 @@ set(XPU_API_LIB_NAME "libxpuapi.so")
 set(XPU_RT_LIB_NAME "libxpurt.so")
 set(XPU_XFT_LIB_NAME "libxft.so")
 
-set(XPU_BASE_DATE "20230323")
+set(XPU_BASE_DATE "20230408")
 set(XPU_XCCL_BASE_VERSION "1.0.13")
 set(XPU_XFT_BASE_VERSION "latest")
 
diff --git a/paddle/phi/kernels/xpu/pool_grad_kernel.cc b/paddle/phi/kernels/xpu/pool_grad_kernel.cc
index afc0bb8fbe12b308d1713fb4dc4507943c5d77dc..1e412e66e68f96466b993ab87d4162320b58e317 100644
--- a/paddle/phi/kernels/xpu/pool_grad_kernel.cc
+++ b/paddle/phi/kernels/xpu/pool_grad_kernel.cc
@@ -114,8 +114,6 @@ void Pool2dGradKernel(const Context& ctx,
     } else if (pooling_type == "avg") {
       // When output dim is 1 * 1 (1 * 1 * 1 in pool_3d), use scale
       // and broadcast kernels to get same output, but better performance.
-      // Since the dim is special in particular models,
-      // use 'export XPU_POOLING_GRAD_SPECIAL=1' to open this path
       if (out_h == 1 && out_w == 1 && std::is_same<T, float>::value) {
         xpu::ctx_guard RAII_GUARD(ctx.x_context());
         float scale = 1.0 / (in_h * in_w);