Data Transfer Problems at the Host-Computer Level and Methods to Improve the Performance of Computations on Heterogeneous Systems

摘要

In theory, graphics processing units largely exceed central processing units by the degree of performance. But in order to load the GPU with computing tasks to the boundary levels of its maximum performance, the task has to be divided into streams. The number of streams should be comparable to or, to some extent, exceed the value of stream processors of the GPU. Nowadays, modern graphics cards have dozens of thousands of streaming processors. This means that the task of ensuring the division into streams in computation of a sufficient degree of the GPU utilization can be either impossible or very difficult, and the solution to this problem is a "weak point" of the whole method of using the GPU, and offsets its advantages in comparison with the CPU. This scientific article is devoted to this problem.