ABSTRACT
Today, customers of very large HPC systems see that the electricity to run the system costs more than the purchase price of the hardware over the life of the massively parallel processor (MPP) systems. Fewer HPC sites can aord to supply the power required by some Petascale systems. Indeed, the biggest challenge to future Etascale systems is the amount of power required to run the system. Today all the chip manufacturers are supplying energy-ecient versions of their top performance chips. e Blue Gene® series of HPC systems are built upon a concept of using many less powerful processors that are more energy ecient. An evolving measurement of HPC systems is FLOPS/WATT; that is, the peak oating point operations supplied by an HPC system divided by the power required to run the system. e approach of using many less power-hungry processors does introduce more complexity in the scaling of applications across the machine. If there are more less-powerful processors, an application must use many more to achieve a performance as good as the previous generation system. Millions of processors in the next-generation HPC system are not inconceivable.
