Plenary lecture 2: high-performance hybrid computing systems and their application in science and engineering
Boris Nikolaevich Chetverushkin · Annual Conference on Computers · 2011
High-performance multiprocessor computing systems are widely applied to the solution of many complicated scientific and engineering problems. However, one of the most significant problem associated with employment of the 1 PTFLOPS systems is their high cost and high power consumption. Therefore, the high-performance multiprocessor computing systems with hybrid architecture, using graphical processing units for acceleration, acquire much wider application. Such supercomputer with the specified architecture by 100 TFLOPS power has been created and installed in the Keldysh Institute of Applied Mathematics of RAS (Moscow). It has relatively low cost? near to $2 million (together with the cooling system and UPS) and low power consumption? near to 70-80 kW. Computing systems based on the graphic cards have great potential for high performance computation. However, the realization of these opportunities is primarily determined by availability of software and by existence of computational algorithms that are compatible with the architecture of the graphic cards. As an example of the hybrid supercomputing system application, the problems of radiation transfer modeling and hydro-and gas dynamics modeling are considered. The gained experience has shown that the hybrid computing systems possessing the high peak performance under skillful adaptation of algorithms to their architecture can be successfully applied to effective modeling of many important scientific and engineering problems.