Enhancing HPC User Support with Generative AI and RAG: Operational Experience on Fugaku User Support
Wednesday, June 24, 2026 3:45 PM to 5:15 PM · 1 hr. 30 min. (Europe/Berlin)
Foyer D-G - 2nd Floor
Research Poster
AI Applications powered by HPC TechnologiesLarge Language Models and Generative AI in HPC
Information
Poster is on display and will be presented at the poster pitch session.
High-performance computing (HPC) systems are becoming increasingly complex, while the diversity and scale of their user communities continue to grow. As a result, user support operations face mounting challenges, including fragmented technical documentation, heterogeneous formats of manuals and FAQs, and a growing volume of user inquiries that are difficult to resolve through conventional keyword-based search systems. These issues lead to significant human workload for support staff and slow problem resolution for users.
This poster presents the design, deployment, and evaluation of a generative AI–based user support system for the supercomputer Fugaku, one of the world’s leading HPC systems. We introduce a retrieval-augmented generation (RAG) chatbot assistant named “AskDona,” which integrates large language models with curated operational knowledge sources. The system was publicly released to Fugaku users in July 2024 after approximately six months of evaluation and proof-of-concept testing across multiple commercial solutions.
AskDona is harnessed with the continuously expanding knowledge base, including system usage guides, programming manuals, job operation documents, optimization case studies for the A64FX processor, FAQ articles from the Fugaku support site, tutorial materials, and other useful and trustworthy information source regarding Fugaku. Through natural language interaction, users can obtain context-aware, concise, and accurate answers taking the advantage of generative AI. This process goes beyond the traditional document search practice as it decomposes the question into pieces and manages the AI agents obtaining the related information from the relevant information source, finally composing the answer to the user. The net effect is the semantic understanding of the user intent and providing the right answer obtained from the trustful information source.
We report quantitative and qualitative operational outcomes observed after the deployment of AskDona. Ticket analysis from the Fugaku support system shows a significant reduction in inquiry-type tickets, with up to a 60% year-over-year decrease in certain periods, despite a steady increase in the number of active users. User interaction logs indicate growing adoption, repeated usage, and effective feedback loops that support continuous system improvement. Additional features, such as adjustable response modes (Fast, Standard, Boost), enable flexible trade-offs between response speed and reasoning depth.
The results demonstrate that RAG-based generative AI can substantially enhance scalability and efficiency in HPC user support while empowering users to resolve issues independently. This work highlights a practical pathway toward evolving AI systems from simple conversational tools into reliable problem-solving partners within large-scale HPC environments.
Contributors:
High-performance computing (HPC) systems are becoming increasingly complex, while the diversity and scale of their user communities continue to grow. As a result, user support operations face mounting challenges, including fragmented technical documentation, heterogeneous formats of manuals and FAQs, and a growing volume of user inquiries that are difficult to resolve through conventional keyword-based search systems. These issues lead to significant human workload for support staff and slow problem resolution for users.
This poster presents the design, deployment, and evaluation of a generative AI–based user support system for the supercomputer Fugaku, one of the world’s leading HPC systems. We introduce a retrieval-augmented generation (RAG) chatbot assistant named “AskDona,” which integrates large language models with curated operational knowledge sources. The system was publicly released to Fugaku users in July 2024 after approximately six months of evaluation and proof-of-concept testing across multiple commercial solutions.
AskDona is harnessed with the continuously expanding knowledge base, including system usage guides, programming manuals, job operation documents, optimization case studies for the A64FX processor, FAQ articles from the Fugaku support site, tutorial materials, and other useful and trustworthy information source regarding Fugaku. Through natural language interaction, users can obtain context-aware, concise, and accurate answers taking the advantage of generative AI. This process goes beyond the traditional document search practice as it decomposes the question into pieces and manages the AI agents obtaining the related information from the relevant information source, finally composing the answer to the user. The net effect is the semantic understanding of the user intent and providing the right answer obtained from the trustful information source.
We report quantitative and qualitative operational outcomes observed after the deployment of AskDona. Ticket analysis from the Fugaku support system shows a significant reduction in inquiry-type tickets, with up to a 60% year-over-year decrease in certain periods, despite a steady increase in the number of active users. User interaction logs indicate growing adoption, repeated usage, and effective feedback loops that support continuous system improvement. Additional features, such as adjustable response modes (Fast, Standard, Boost), enable flexible trade-offs between response speed and reasoning depth.
The results demonstrate that RAG-based generative AI can substantially enhance scalability and efficiency in HPC user support while empowering users to resolve issues independently. This work highlights a practical pathway toward evolving AI systems from simple conversational tools into reliable problem-solving partners within large-scale HPC environments.
Contributors:
Format
on-demandon-site