diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5ca15da18..a44463b8d 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -193,6 +193,11 @@ if(AER_THRUST_SUPPORTED)
 		endif()
 		cuda_select_nvcc_arch_flags(AER_CUDA_ARCH_FLAGS ${AER_CUDA_ARCH})
 		set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -ccbin ${CMAKE_CXX_COMPILER} ${AER_CUDA_ARCH_FLAGS} -DAER_THRUST_CUDA -std=c++14 -I${AER_SIMULATOR_CPP_SRC_DIR} -isystem ${AER_SIMULATOR_CPP_SRC_DIR}/third-party/headers -use_fast_math --expt-extended-lambda")
+		# We have to set SIMD flags globally because there seems to be a bug in FindCUDA which doesn't allow us to set per-file compilation flags.
+		# The implications is that users downloading the PyPi wheel package for the GPU, need to have a CPU with AVX2 support otherwise
+		# Aer will crash with: "Unknow instruction" exception.
+		# This will be fixed here: https://github.com/Qiskit/qiskit-aer/
+		set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS}; --compiler-options;-mfma,-mavx2")
 		set(AER_COMPILER_DEFINITIONS ${AER_COMPILER_DEFINITIONS} THRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_CUDA)
 		set(THRUST_DEPENDANT_LIBS "")
 	elseif(AER_THRUST_BACKEND STREQUAL "TBB")