It seems they are finally enable the 16bit float extension in their OpenCL implementation.