Arithmetic for relative accuracy

Of the three factors named in Moore's first Law that drive the advance of computational systems, circuit design receives relatively little mention. We introduce here a circuit variety that allows to include accuracy considerations. It is shown that accuracy-drive can be effectively realised and leads to 60% speed improvement. Details are given of a floating-point unit with full hardware support of complex calculations, specifically tailored to speed-up MD-simulations on the GROMACS scientific parallel computer.<>