So does that mean although the type is bfloat16 but the computation is actually run in fp32 way in GPU?