diff --git a/paddle/cuda/include/hl_matrix_type.cuh b/paddle/cuda/include/hl_matrix_type.cuh
index 77f73167fe6ea2e301b3f68e51d32f3e1578ec93..12c717b612da907ac7aedf3e7787bdbb96948be2 100644
--- a/paddle/cuda/include/hl_matrix_type.cuh
+++ b/paddle/cuda/include/hl_matrix_type.cuh
@@ -39,6 +39,8 @@ typedef double2 vecType;
 #elif defined(__SSE3__)
 #include "hl_cpu_simd_sse.cuh"
 #elif (defined(__ARM_NEON) || defined(__ARM_NEON__)) && !defined(__NVCC__)
+// Currently nvcc does not support neon intrinsic.
+// TODO: Extract simd intrinsic implementation from .cu files.
 #include "hl_cpu_simd_neon.cuh"
 #else
 #include "hl_cpu_scalar.cuh"