Introducing Replicated VLSI to Supercomputing: the FPS-164/MAX Scientific Computer

Large-scale scientific computing is dominated by the "crunching of numbers." By performing billions of multiplies and adds on millions of simulated data points, scientists can model physical phenomena, like the flow of air around a supersonic airplane or the movement of storms in the atmosphere. The data values range over more than fifty orders of magnitude, from the inside of a quark to the circumference of the universe. Because a simulation run is made up of thousands of iterated computation steps, at least eight significant digits must be retained by each arithmetic operation to avoid excessive round-off errors in the results. The 16and 32-bit integer arithmetic common in business computing does not provide adequate range or accuracy for scientific use. Instead, scientific applications normally use 64-bit floating-point arithmetic. IEEE standard double-precision arithmetic, I for example, provides a dynamic range of 616 orders ofmagnitude and a precision of 15 significant digits. Because arithmetic speed usually dominates overall execution times, the performance of scientific computers is traditionally measured in MFLOPs, or millions of floating-point operations per second. Current processor implementations range from a speed ofone twentieth MFLOP for the best personal computer CPU to several hundred MFLOPs per supercomputer CPU. VLSI parts for number crunching