diff --git a/paddle/fluid/operators/gelu_op.cu b/paddle/fluid/operators/gelu_op.cu
index da7ed05ddf55c94c9011e24e95b9c546a5173b17..17f74464a811c5c09411cbbbd4c2e24b1948ee12 100644
--- a/paddle/fluid/operators/gelu_op.cu
+++ b/paddle/fluid/operators/gelu_op.cu
@@ -24,7 +24,7 @@ namespace operators {
 #ifdef __NVCC__
 template <bool FastMode>
 static __device__ __forceinline__ float FP32FastTanh(float x) {
-#if __CUDA_ARCH__ >= 750 && !defined(_WIN32)
+#if __CUDA_ARCH__ >= 750 && CUDA_VERSION >= 11000 && !defined(_WIN32)
   if (FastMode) {
     float y;
     asm("tanh.approx.f32 %0,%1; \n\t" : "=f"(y) : "f"(x));