Implications of Reduced-Precision Computations in HPC: Performance, Energy and Error