Hardware and software for distributed and supercomputer systems

Research Article

DOI 10.25209/2079-3316-2026-17-3-237-273

Backpropagation algorithm in the dataflow paradigm

Dmitry Nikolayevich Zmejev¹, Nikolay Nikolayevich Levchenko², Arkady Valentinovich Klimov³

1-3	National Research Center „Kurchatov Institute“, Moscow, Russia
1	zmejevdn@list.ru
2	n2lev@yandex.ru
3	arkady.klimov@gmail.com

Abstract. The article discusses the issue of developing and implementing backpropagation algorithm in the dataflow paradigm. The principles of dataflow computing differ significantly from traditional control-flow computing, just as the dataflow programming paradigm differs from the imperative one.
Programs created in the dataflow paradigm are intrinsically parallel, since the task parallelism is extracted automatically from the data flow at the hardware level. Programs created in the dataflow paradigm are initially parallel. The article provides a detailed description of the backpropagation dataflow algorithm and the program that runs on the parallel dataflow computing system (PDCS) «Buran».
The dataflow program is compact and versatile in its code. It automatically scales to the entire system, and its program code does not contain references to library functions and is based solely on basic arithmetic operations such as addition, multiplication, and comparison. The program is capable of training perceptrons of any dimension without modification and recompilation of the program code. The size and structure of the trainable perceptron is determined by the initial data.
The experimental part of the article presents the results of PDCS behavior study when executing the backpropagation program, analyzes the effect of the trainable batch size on the overall efficiency of the dataflow program, and evaluates the use of various methods of hardware distribution of computations. In addition, the potential of using the PDCS for parallel execution of several training tasks at the same time is considered. (In Russian).

Keywords: backpropagation algorithm, parallel programming, dataflow computing model, dataflow programming paradigm, parallel dataflow computing system

MSC-2020

68Q09; 68T01, 68W10

Acknowledgments: The work was carried out within the state assignment of NRC «Kurchatov institute»

For citation: Dmitry N. Zmejev, Nikolay N. Levchenko, Arkady V. Klimov. Backpropagation algorithm in the dataflow paradigm. Program Systems: Theory and Applications, 2026, 17:3, pp. 237–273. (In Russ.). https://psta.psiras.ru/2026/3_237-273.

Full text of article (PDF): https://psta.psiras.ru/read/psta2026_3_237-273.pdf.

The article was submitted 30.06.2026; approved after reviewing 30.07.2026; accepted for publication 31.07.2026; published online 18.09.2026.