Abstract:
Specialized hardware solutions play an important role in the
development of supercomputer technologies. Currently, most computing systems
of maximum performance use mathematical coprocessors of various types. So the
development of applied software solutions designed to realize the potential of
modern computing platforms requires ensuring the efficient use of hardware
accelerators. The course of work on a software system for simulating the behavior
of charge carriers in graphene needs to solve the problem of supporting such
accelerators and to investigate the efficiency of the solution obtained. The
current situation and the prospects for the next few years suggest the NVIDIA
accelerators and CUDA software technology, but the hardware architecture of
NVIDIA accelerators is fundamentally different from the CPU architecture, and
therefore the mathematical libraries adapted for CUDA do not support the entire
range of algorithms used in the original version of the program. The paper
presents the features of an implementation of CUDA support and the results of
comparative testing of the obtained solution using the example of a problem with
realistic characteristics.
Key words and phrases:high performance computing, hybrid architectures, CUDA, graphene,
quantum kinetic equation.
Fulltext:
PDF file (1257 kB)