<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Publishing DTD v1.3 20210610//EN" "JATS-journalpublishing1-3.dtd">
<article article-type="research-article" dtd-version="1.3" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xml:lang="ru"><front><journal-meta><journal-id journal-id-type="publisher-id">sapi</journal-id><journal-title-group><journal-title xml:lang="ru">Системный анализ и прикладная информатика</journal-title><trans-title-group xml:lang="en"><trans-title>«System analysis and applied information science»</trans-title></trans-title-group></journal-title-group><issn pub-type="ppub">2309-4923</issn><issn pub-type="epub">2414-0481</issn><publisher><publisher-name>Belarusian National Technical University</publisher-name></publisher></journal-meta><article-meta><article-id pub-id-type="doi">10.21122/2309-4923-2026-1-54-59</article-id><article-id custom-type="elpub" pub-id-type="custom">sapi-799</article-id><article-categories><subj-group subj-group-type="heading"><subject>Research Article</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="ru"><subject>Обработка информации и принятие решений</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="en"><subject>Data processing and decision–making</subject></subj-group></article-categories><title-group><article-title>Стратегии параллелизма как ключевой фактор развертывания Large Language Models на базе потребительских GPU</article-title><trans-title-group xml:lang="en"><trans-title>Parallelism strategies as a key factor for deploying Large Language Models on consumer gpus</trans-title></trans-title-group></title-group><contrib-group><contrib contrib-type="author" corresp="yes"><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Курочка</surname><given-names>К. С.</given-names></name><name name-style="western" xml:lang="en"><surname>Kurochka</surname><given-names>K. S.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Курочка Константин Сергеевич − кандидат технических наук, доцент.</p><p>г. Гомельkurochka@gstu.by</p><p> </p></bio><bio xml:lang="en"><p>Konstantin S. Kurochka – PhD in Engineering, Associate Professor.</p><p>Gomelkurochka@gstu.by</p></bio><email xlink:type="simple">kurochka@gstu.by</email><xref ref-type="aff" rid="aff-1"/></contrib><contrib contrib-type="author" corresp="yes"><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Башаримов</surname><given-names>Ю. С.</given-names></name><name name-style="western" xml:lang="en"><surname>Basharymau</surname><given-names>Yu. S.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Башаримов Юрий Сергеевич − ассистент кафедры.</p><p>г. Гомельbasharymauyury@gmail.com</p></bio><bio xml:lang="en"><p>Yury S. Basharymau – Assistant.Gomelbasharymauyury@gmail.com</p></bio><email xlink:type="simple">basharymauyury@gmail.com</email><xref ref-type="aff" rid="aff-1"/></contrib><contrib contrib-type="author" corresp="yes"><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Ёвженко</surname><given-names>Ю. Д.</given-names></name><name name-style="western" xml:lang="en"><surname>Youzhanka</surname><given-names>Yu. D.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Ёвженко Юрий Дмитриевич − магистрант.</p><p>г. Гомельyuevzhenko@gmail.com</p></bio><bio xml:lang="en"><p>Yury D. Youzhanka – Master’s Student.</p><p>Gomelyuevzhenko@gmail.com</p></bio><email xlink:type="simple">yuevzhenko@gmail.com</email><xref ref-type="aff" rid="aff-1"/></contrib></contrib-group><aff-alternatives id="aff-1"><aff xml:lang="ru"><institution>Гомельский государственный технический университет им. П. О. Сухого</institution><country>Беларусь</country></aff><aff xml:lang="en"><institution>Sukhoi State Technical University of Gomel</institution><country>Belarus</country></aff></aff-alternatives><pub-date pub-type="collection"><year>2026</year></pub-date><pub-date pub-type="epub"><day>10</day><month>04</month><year>2026</year></pub-date><volume>0</volume><issue>1</issue><fpage>54</fpage><lpage>59</lpage><permissions><copyright-statement>Copyright &amp;#x00A9; Курочка К.С., Башаримов Ю.С., Ёвженко Ю.Д., 2026</copyright-statement><copyright-year>2026</copyright-year><copyright-holder xml:lang="ru">Курочка К.С., Башаримов Ю.С., Ёвженко Ю.Д.</copyright-holder><copyright-holder xml:lang="en">Kurochka K.S., Basharymau Y.S., Youzhanka Y.D.</copyright-holder><license xml:lang="ru" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>Данная работа распространяется под лицензией Creative Commons Attribution 4.0.</license-p></license><license xml:lang="en" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>This work is licensed under a Creative Commons Attribution 4.0 License.</license-p></license></permissions><self-uri xlink:href="https://sapi.bntu.by/jour/article/view/799">https://sapi.bntu.by/jour/article/view/799</self-uri><abstract><p>Экспоненциальный рост размеров больших языковых моделей (LLM) создает существенные барьеры для их локального развертывания, обусловленные нехваткой видеопамяти (VRAM) на одиночных устройствах. Целью работы является выявление и обоснование наиболее эффективной стратегии параллелизма для инференса LLM на кластерах из потребительских графических процессоров (GPU), объединенных медленной шиной PCIe. Методы исследования включали проведение серии вычислительных экспериментов для сравнения монолитной архитектуры (NVIDIA RTX A6000) и распределенной системы (2x NVIDIA RTX 3090) с использованием фреймворка vLLM. Анализировалось влияние тензорного (Tensor Parallelism) и конвейерного (Pipeline Parallelism) параллелизма на ключевые метрики: пропускную способность, задержку (TTFT, TPOT) и стабильность энергопотребления при запуске модели DeepSeek-R1-DistillLlama-14B. Результаты однозначно указывают на непригодность тензорного параллелизма для систем без NVLink из-за критических задержек синхронизации. Доказано, что конвейерный параллелизм является единственной жизнеспособной стратегией для PCIe-кластеров, обеспечивая высокую пропускную способность, несмотря на наличие периодов простоя («пузырей») и менее стабильный профиль энергопотребления по сравнению с монолитным решением. В заключении сформулированы рекомендации по использованию мульти-GPU конфигураций: они являются оптимальным экономическим выбором для задач, критичных к объему памяти, таких как Retrieval-Augmented Generation (RAG), позволяя масштабировать VRAM значительно дешевле профессиональных аналогов.</p></abstract><trans-abstract xml:lang="en"><p>The exponential growth in the size of Large Language Models (LLMs) creates significant barriers to their local deployment, primarily due to Video RAM (VRAM) shortages on single devices. The aim of this work is to identify and substantiate the most effective parallelism strategy for LLM inference on consumer Graphics Processing Unit (GPU) clusters connected via a slow PCIe bus. Research methods included a series of experiments comparing a monolithic architecture (NVIDIA RTX A6000) and a distributed system (2x NVIDIA RTX 3090) using the vLLM framework. The impact of Tensor Parallelism (TP) and Pipeline Parallelism (PP) on key metrics – throughput, latency (TTFT, TPOT), and power consumption stability – was analyzed while running the DeepSeek-R1-Distill-Llama-14B model. The results unequivocally indicate the unsuitability of Tensor Parallelism for systems without NVLink due to critical synchronization delays. It is proven that Pipeline Parallelism is the only viable strategy for PCIe clusters, ensuring high throughput despite the presence of idle periods («bubbles») and a less stable power consumption profile compared to the monolithic solution. In conclusion, recommendations for using multi-GPU configurations are formulated: they represent the optimal economic choice for memory-critical tasks, such as Retrieval-Augmented Generation (RAG), allowing VRAM scaling at a significantly lower cost than professional analogs.</p></trans-abstract><kwd-group xml:lang="ru"><kwd>большие языковые модели</kwd><kwd>LLM</kwd><kwd>инференс</kwd><kwd>тензорный параллелизм</kwd><kwd>конвейерный параллелизм</kwd><kwd>vLLM</kwd><kwd>GPU</kwd><kwd>CUDA</kwd><kwd>NVLink</kwd><kwd>PCIe</kwd><kwd>RAG</kwd></kwd-group><kwd-group xml:lang="en"><kwd>Large Language Models</kwd><kwd>LLM</kwd><kwd>inference</kwd><kwd>Tensor Parallelism</kwd><kwd>Pipeline Parallelism</kwd><kwd>vLLM</kwd><kwd>GPU</kwd><kwd>CUDA</kwd><kwd>NVLink</kwd><kwd>PCIe</kwd><kwd>RAG</kwd></kwd-group></article-meta></front><back><ref-list><title>References</title><ref id="cit1"><label>1</label><citation-alternatives><mixed-citation xml:lang="ru">Attention is all you need / A. Vaswani, N. Shazeer, N. Parmar [et al.] // Advances in Neural Information Processing Systems. 2017. Vol. 30. DOI:10.48550/arXiv.1706.03762.</mixed-citation><mixed-citation xml:lang="en">Vaswani A., Shazeer N., Parmar N., Uszkoreit J., Jones L., Gomez A.N., et al. Attention is all you need. Advances in Neural Information Processing Systems. 2017;30. https://doi.org/10.48550/arXiv.1706.03762.</mixed-citation></citation-alternatives></ref><ref id="cit2"><label>2</label><citation-alternatives><mixed-citation xml:lang="ru">Курочка, К. С. Технология трансляции естественно-языковых правил мерчандайзинга в цифровые планограммы / К. С. Курочка, Ю. Д. Ёвженко // Информатика. 2025. Т. 22, № 4. С. 55–64. DOI: 10.37661/1816-0301-2025-22-4-55-64.</mixed-citation><mixed-citation xml:lang="en">Kurochka K.S., Youzhanka Y.D. The technology of translating natural language merchandising rules into digital planograms. Informatics. 2025;22(4):55−64 (in Russian). http://dx.doi.org/10.37661/1816-0301-2025-22-4-55-64.</mixed-citation></citation-alternatives></ref><ref id="cit3"><label>3</label><citation-alternatives><mixed-citation xml:lang="ru">Scaling laws for neural language models / J. Kaplan, S. McCandlish, T. Henighan [et al.] // arXiv preprint arXiv:2001.08361. 2020. DOI: 10.48550/arXiv.2001.08361.</mixed-citation><mixed-citation xml:lang="en">Kaplan J., McCandlish S., Henighan T., Brown T.B., Chess B., Child R., et al. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361. 2020. https://doi.org/10.48550/arXiv.2001.08361.</mixed-citation></citation-alternatives></ref><ref id="cit4"><label>4</label><citation-alternatives><mixed-citation xml:lang="ru">Large Language Models: A survey / S. Minaee, T. Mikolov, N. Nikzad [et al.] // arXiv preprint arXiv:2402.06196. 2024. DOI: 10.48550/arXiv.2402.06196.</mixed-citation><mixed-citation xml:lang="en">Minaee S., Mikolov T., Nikzad N., Chenaghlu M., Socher R., Amatriain X., et al. Large Language Models: A survey. arXiv preprint arXiv:2402.06196. 2024. https://doi.org/10.48550/arXiv.2402.06196.</mixed-citation></citation-alternatives></ref><ref id="cit5"><label>5</label><citation-alternatives><mixed-citation xml:lang="ru">Курочка, К. С. Нейросетевая модель автоматической генерации тестовых заданий для студентов в системе Moodle на основе анализа методических материалов / К. С. Курочка, Ю. С. Башаримов // Цифровая трансформация. 2025. Т. 31, № 3. С. 66–75. DOI: 10.35596/1729-7648-2025-31-3-66-75.</mixed-citation><mixed-citation xml:lang="en">Kurochka K.S., Basharymau Y.S. Neural network model for automated test generation for students in the Moodle system based on the analysis of methodological materials. Digital Transformation. 2025;31(3):66–75 (in Russian). http://dx.doi.org/10.35596/1729-7648-2025-31-3-66-75.</mixed-citation></citation-alternatives></ref><ref id="cit6"><label>6</label><citation-alternatives><mixed-citation xml:lang="ru">Megatron-LM: Training multi-billion parameter language models using model parallelism / M. Shoeybi, M. Patwary, R. Puri [et al.] //arXiv preprint arXiv:1909.08053. 2019. DOI: 10.48550/arXiv.1909.08053.</mixed-citation><mixed-citation xml:lang="en">Shoeybi M., Patwary M., Puri R., Le Gresley P., Casper J., Catanzaro B. Megatron-LM: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053. 2019. https://doi.org/10.48550/arXiv.1909.08053.</mixed-citation></citation-alternatives></ref><ref id="cit7"><label>7</label><citation-alternatives><mixed-citation xml:lang="ru">Efficiently Scaling Transformer Inference / R. Pope, S. Douglas, A. Chowdhery [et al.] // Proceedings of Machine Learning and Systems (MLSys 2023). 2023. Vol. 5. P. 606–624. URL: https://proceedings.mlsys.org/paper_files/paper/2023/hash/c4be71ab8d24cdfb45e3d06dbfca2780-Abstract-mlsys2023.html (date of access: 10.02.2026).</mixed-citation><mixed-citation xml:lang="en">Pope R., Douglas S., Chowdhery A., Devlin J., Bradbury J., Heek J., et al. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems (MLSys 2023). 2023;(5):606–624. Available at: https://proceedings.mlsys.org/paper_files/paper/2023/hash/c4be71ab8d24cdfb45e3d06dbfca2780-Abstract-mlsys2023.html (accessed 10 February 2026).</mixed-citation></citation-alternatives></ref><ref id="cit8"><label>8</label><citation-alternatives><mixed-citation xml:lang="ru">PowerInfer: Fast Large Language Model serving with a consumer-grade GPU / Y. Song, Z. Mi, H. Xie, H. Chen // Proceedings of the ACM SIGOPS 30th Symposium on Operating Systems Principles (SOSP’24). 2024. P. 590-606. DOI: 10.1145/3694715.3695964.</mixed-citation><mixed-citation xml:lang="en">Song Y., Mi Z., Xie H., Chen H. Powerinfer: Fast Large Language Model serving with a consumer-grade GPU. Proceedings of the ACM SIGOPS 30th Symposium on Operating Systems Principles (SOSP’24). 2024:590–606. https://doi.org/10.1145/3694715.3695964.</mixed-citation></citation-alternatives></ref><ref id="cit9"><label>9</label><citation-alternatives><mixed-citation xml:lang="ru">GPipe: Efficient training of giant neural networks using pipeline parallelism / Y. Huang, Y. Cheng, A. Bapna [et al.] // Advances in Neural Information Processing Systems (NeurIPS 2019). 2019. Vol. 32. URL: https://proceedings.neurips.cc/papers/search?q=GPipe%3A+Efficient+training+of+giant+neural+networks+using+pipeline+parallelism+ (date of access: 10.02.2026).</mixed-citation><mixed-citation xml:lang="en">Huang Y., Cheng Y., Bapna A., Firat O., Chen D., Chen M., et al. GPipe: Efficient training of giant neural networks using pipeline parallelism. Advances in Neural Information Processing Systems (NeurIPS 2019). 2019;32. Available at: https://proceedings.neurips.cc/papers/search?q=GPipe%3A+Efficient+training+of+giant+neural+networks+using+pipeline+parallelism+ (accessed 10 February 2026).</mixed-citation></citation-alternatives></ref><ref id="cit10"><label>10</label><citation-alternatives><mixed-citation xml:lang="ru">Efficient memory management for Large Language Model serving with PagedAttention / W. Kwon, Z. Li, S. Zhuang [et al.] // Proceedings of the 29th Symposium on Operating Systems Principles (SOSP’23). 2023. P. 611–626. DOI: 10.1145/3600006.3613165.</mixed-citation><mixed-citation xml:lang="en">Kwon W., Li Z., Zhuang S., Sheng Y., Zheng L., Yu C.H., et al. Efficient memory management for Large Language Model serving with PagedAttention. Proceedings of the 29th Symposium on Operating Systems Principles (SOSP’23). 2023:611–626. https://doi.org/10.1145/3600006.3613165.</mixed-citation></citation-alternatives></ref><ref id="cit11"><label>11</label><citation-alternatives><mixed-citation xml:lang="ru">DeepSeek-V2: A strong, economical, and efficient Mixture-of-Experts language model / A. Liu, B. Feng, B. Wang [et al.] // arXiv preprint arXiv:2405.04434. 2024. DOI: 10.48550/arXiv.2405.04434.</mixed-citation><mixed-citation xml:lang="en">Liu A., Feng B., Wang B., Wang B., Liu B., Zhao C., et al. DeepSeek-V2: A strong, economical, and efficient Mixture-of-Experts language model. arXiv preprint arXiv:2405.04434. 2024. https://doi.org/10.48550/arXiv.2405.04434.</mixed-citation></citation-alternatives></ref><ref id="cit12"><label>12</label><citation-alternatives><mixed-citation xml:lang="ru">Retrieval-augmented generation for knowledge-intensive NLP tasks / P. Lewis, E. Perez, A. Piktus [et al.] // Advances in Neural Information Processing Systems. (NeurIPS 2020). 2020. Vol. 33. URL: https://proceedings.neurips.cc/papers/search?q=Retrieval-Augmented+Generation+for+Knowledge-Intensive+NLP+Tasks+ (date of access: 10.02.2026).</mixed-citation><mixed-citation xml:lang="en">Lewis P., Perez E., PiktusA., Petroni F., KarpukhinV., Goyal N., et al. Retrieval-augmented generation for knowledgeintensive NLP tasks. Advances in Neural Information Processing Systems (NeurIPS 2020). 2020;33. Available at:https://proceedings.neurips.cc/papers/search?q=Retrieval-Augmented+Generation+for+Knowledge-Intensive+NLP+Tasks+ (accessed 10 February 2026).</mixed-citation></citation-alternatives></ref></ref-list><fn-group><fn fn-type="conflict"><p>The authors declare that there are no conflicts of interest present.</p></fn></fn-group></back></article>
