diff --git a/Eigen/Core b/Eigen/Core
index 3edbe6585..834ff9415 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -200,7 +200,10 @@
 #if defined __CUDACC__
   #define EIGEN_VECTORIZE_CUDA
   #include <vector_types.h>
-  #include <cuda_fp16.h>
+  #if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+    #define EIGEN_HAS_CUDA_FP16
+    #include <cuda_fp16.h>
+  #endif
 #endif
 
 #if (defined _OPENMP) && (!defined EIGEN_DONT_PARALLELIZE)
diff --git a/Eigen/src/Core/arch/CUDA/TypeCasting.h b/Eigen/src/Core/arch/CUDA/TypeCasting.h
index a8c06ff48..279fd4fd0 100644
--- a/Eigen/src/Core/arch/CUDA/TypeCasting.h
+++ b/Eigen/src/Core/arch/CUDA/TypeCasting.h
@@ -14,6 +14,8 @@ namespace Eigen {
 
 namespace internal {
 
+#if defined(EIGEN_HAS_CUDA_FP16)
+
 template<>
 struct scalar_cast_op<float, half> {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
@@ -92,6 +94,7 @@ template<> EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) {
 #endif
 }
 
+#endif
 
 } // end namespace internal
 
diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
index e9f5dd968..aee222a14 100644
--- a/unsupported/test/cxx11_tensor_of_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
@@ -19,6 +19,7 @@
 
 using Eigen::Tensor;
 
+#ifdef EIGEN_HAS_CUDA_FP16
 void test_cuda_conversion() {
   Eigen::CudaStreamDevice stream;
   Eigen::GpuDevice gpu_device(&stream);
@@ -52,9 +53,11 @@ void test_cuda_conversion() {
   gpu_device.deallocate(d_half);
   gpu_device.deallocate(d_conv);
 }
-
+#endif
 
 void test_cxx11_tensor_of_float16_cuda()
 {
+#ifdef EIGEN_HAS_CUDA_FP16
   CALL_SUBTEST_1(test_cuda_conversion());
+#endif
 }