The explosive growth in the number of created processors for working with neural networks (neuroprocessors) and other samples of the electronic component base (ECB) for artificial intelligence systems (AI) is proposed to be considered as a new direction in the development of ECB supercomputer technologies (HPC tech), for which an even greater degree. It is important to overcome the problems of working with memory, energy efficiency, increasing the parallelism of operations and reducing the overhead of communication. A warning is made that in the work on AI technologies in general, it is necessary to take into account the errors in organizing work on domestic HPC, since skewing was allowed in this 20-year process – most attention was paid to everything except the development of the domestic electronic components . As a result, a severe lag formed behind the foreign ECB HPC and, as a result, there were significant modern problems of import substitution in this area. Work on domestic AI technologies, as well as on HPC technologies, can be divided into two areas: the first is the adaptation and development of foreign technologies, the development of original algorithms and software; the second is the development of its own electronic components. At the same time, the work on electronic components of AI, according to the authors, is of great importance, which is explained by the future mass character and variety of the market of AI systems, but with the predicted halt in the development of microelectronic CMOS (complementary metal-oxidesemiconductor) technologies, while the industrial development of new ones (“post Moore”) technology is expected only at the end of the coming decade. In the development of domestic HPC, a special role was played by the target programs of the Union State (CPSG) SKIF and TRIAD. For the establishment of domestic AI technologies, at least a Russian federal target program, and preferably a CPSG, is also required. ECB samples created in the framework of such programs should support not only deep learning neural networks (DNN) and neuromorphic (spike) networks (SNN), but also other basic AI methods. Taking into account the history of the organization and development of work on domestic HPC technology, as well as their current state, the article discusses the organization of AI technology programs with a focus on the development of electronic components of AI.

AI hardware, neuroprocessors, manytiled architecture, multithreaded core, domain specific architecture, deep learning neuro networks, neurumorthic (spaiking) networks, training, inference

Adamov Andrey Anatolievich,  a.adamov@module.ru,  Recearch Center ”Module”

Eisymont Leonid Konstantinovich,  verger-lk@yandex.ru,  Recearch Center ”Module”