Recent Technology Trends in High-Performance Computing: Quantifying the Divergence of Top and the Rest

Exploratory analysis methods were used to study basic characteristics of computing systems from TOP500_Lists. One of the peculiarities of the distribution of computing systems by performance is that it sufficiently well obeys an analog of the empirical Zipf’s law, in which logarithm of performance is reciprocal to the rank of computing system. Based on this observation we can divide all systems from the lists into several performance classes: top, high, base, and entry levels. Our analysis also revealed differences between these classes in other characteristics besides, the computational performance, e. g., such as power consumption. For all performance classes, trends in evolution of the basic characteristics of TOP500 computing systems were described and, where possible, comments were provided to explain their behavior. Performance and energy efficiency of the TOP500_List computing systems in the next 5–10 years were estimated using simple linear models obtained by the least-square method. We have found that energy consumption needed for entry-level supercomputers to surpass the threshold value of performance and to enter into TOP500_List will decrease during this period.