M. V. Zibul

Traditionally, for large amounts of computing in the modern scientific world use powerful computing systems on supercomputers. However, the use of supercomputers has several known drawbacks such as high cost, need for special training and special server room engineering solutions to ensure the work environment, the need for hardware and software only specially trained personnel. Today supercomputer takes up space, consumes a lot of electricity, is rapidly becoming obsolete and expensive to maintain. Exceptional computational power of the GPU due to the peculiarities of its architecture. Unlike the CPU, which consists of multiple cores, a graphics processor was originally designed as a multi-core structure in which the number of nuclei in the hundreds and thousands. The difference in the architecture and makes a difference in how it works. If the CPU architecture involves sequential processing of information, the GPU has historically meant for processing computer graphics, so is designed for massively parallel computing. Each of these two architectures has its advantages. CPU works better with sequential tasks. A high volume of information being processed with the natural parallelism has an obvious advantage GPU. An important condition is that only the existence of such a parallelism in the problem. The advantage of hybrid supercomputer is a price that is comparable to the cost of several professional workstations at a much higher performance hybrid system. It is noted the relative ease of deployment and lower maintenance costs, and greater efficiency of the cooling system and energy consumption. Scope of the present and potential applications of hybrid supercomputing is extensive: - theoretical, experimental and applied computational problems in various fields of science; - industrial problem of calculating, analyzing, and modeling (CAD, oil and gas industry, etc.); - Finance (calculation of credit and insurance risks); - medicine (3D tomography, ultrasound imaging, etc.). A necessary condition for the introduction of hybrid supercomputer is the availability of appropriate hardware and software. However, this condition is not sufficient, there are problems of building large-scale use of graphics processors, especially in the domestic market. Of particular relevance, issues of migration to hybrid supercomputers existing applications with large amounts of code. The problem of programming for hybrid supercomputing is one of the most difficult problems. This complexity is due to the fact that the process of parallel algorithms is difficult to formalize, and in most cases, success depends on the art of parallel programming. One solution to this problem is to use data parallelism. The two major GPU manufacturer, NVIDEA and AMD, have developed and announced a platform for high-performance hybrid called CUDA (Compute Unified Device Architecture) and CTM (Close To Metal or AMD Stream Computing), respectively. These programming models are made with a view to directly access the hardware capabilities of graphics cards.