AI is making high-performance computing more efficient by improving scheduling, performance modeling and fault detection, with scheduling the most active research area; yet integration, MLOps practices and benchmarking standards lag behind. Domain-specific models, including specialized LLMs, often outperform general-purpose alternatives for HPC automation.

Main Finding

AI (broadly defined to include ML and optimization) can materially improve the operational efficiency, reliability, and automation of HPC systems across six application areas—performance estimation, performance optimization, scheduling, surrogate modeling, fault detection, and language-model-based automation. The strongest activity is in scheduling (including RL and hybrid approaches). Supervised performance estimation is foundational, GNNs/time-series models improve anomaly detection, and domain-specialized LMs show promise for coding and operational automation. Realizing these gains requires advances in MLOps, standardized components and benchmarks, and attention to deployment costs and dataset-shift robustness.

Key Points

• Scope and scale

  • Manual review of ~1,800 papers (2019–2025 window); 823 relevant to the HPC/AI intersection; 74 peer-reviewed papers focused on “AI for HPC” retained for detailed analysis (≈9% vs. 91% focused on “HPC for AI”).
  • Papers were grouped into six application areas: performance estimation, performance optimization, scheduling, surrogate models, fault detection, and LLM-based automation.

• Application-area highlights

  • Scheduling: the most active field. Approaches range from research RL schedulers to production-friendly hybrids combining ML estimators with heuristics or constraint solvers.
  • Performance estimation: supervised learning models (job duration, memory, CPU) are core enablers for better scheduling and optimization.
  • Performance optimization: ML and optimization methods (Bayesian opt, ILP, evolutionary algorithms) tune parameters, select libraries, or guide parallelization.
  • Surrogate modeling: ML models replace or approximate expensive simulations to reduce runtime.
  • Fault detection: graph neural networks and time-series models capture spatio-temporal dependencies in telemetry to detect anomalies more reliably.
  • LLM-based automation: domain-specialized LMs outperform general-purpose LLMs on HPC coding/automation tasks and can enable scripting, configuration, and operator assistance.

• Integration & practical considerations

  • Hybrid approaches (ML + heuristics/constraints) are common because heuristics are easier to deploy and maintain; ML offers higher potential gains but brings operational overhead (data pipeline, monitoring, MLOps).
  • Opportunities include LLM-based OS/automation concepts and coupling estimators + optimizers + schedulers into end-to-end stacks.
  • Major needs: MLOps practices adapted to HPC, shared APIs/data formats, standardized benchmarks to compare approaches, and methods to handle dataset shift and model drift.

• Environmental/operational context

  • Accelerator adoption in top HPC systems increased (e.g., from 64/500 in June 2014 to 195/500 in June 2024).
  • Energy efficiency has improved, but total energy consumption and scale of systems have also risen (reported average power from ~1.1 MW to ~2.1 MW over a decade), motivating energy-aware AI solutions.

Data & Methods

• Search strategy

  • Databases: IEEE Xplore, Scopus, Google Scholar, Google web search.
  • Query: conjunction of AI-method terms (“artificial intelligence” OR “machine learning” OR “deep learning” OR “reinforcement learning” OR “LLM”) AND HPC-goal terms (“HPC” OR “software performance” OR “accelerator”).
  • Inclusion: peer-reviewed papers from 2018 onward; focused on AI applied to HPC operations/software rather than “HPC for AI” or chip micro-architecture design.
  • Manual screening and classification (authors found automated classifiers insufficient); cross-validation among co-authors.

• Classification rules and limitations

  • Papers assigned to one of six categories using a “terminal-component” rule (classify by the final AI component in the pipeline).
  • Limitations: manual snapshot review (results may change as literature evolves), exclusion of non–peer-reviewed reports, possible residual classification bias, and deliberate exclusion of pre-2018 works and micro-architecture studies.
  • Quantitative summary: ~1,800 screened → 823 HPC/AI intersection → 74 AI-for-HPC papers retained.

Implications for AI Economics

• Potential economic benefits

  • Lower operational costs: improved scheduling and resource estimation can increase throughput and utilization, reducing queuing delays and wasted cycles.
  • Reduced energy costs and carbon footprint: energy-aware scheduling, surrogate models (reducing full-run simulations), and dynamic optimization can lower power consumption per useful computation—valuable where energy is a major cost center.
  • Productivity gains: domain-specific LMs can speed up code generation, deployment, and operator workflows, reducing labor time per task.
  • Better asset utilization: more accurate job profiling and dynamic allocation can defer capital expenditures by squeezing more useful work from existing nodes.

• Investment trade-offs and costs

  • Up-front and ongoing costs: data collection, model training, MLOps infrastructure, monitoring, and retraining for dataset shift. These can be substantial relative to simple heuristic approaches.
  • Deployment complexity: hybrid solutions (ML + heuristics) often offer the best marginal ROI, combining deployability with gains; pure ML solutions may face higher operational risk.
  • Model maintenance risk: model drift and the need for robust validation increase recurring costs and governance requirements.
  • Compute budget: training/serving models (especially LLMs or large GNNs) consumes compute; net economic gain depends on whether model-driven improvements exceed these added costs.

• Market and policy implications

  • Product demand: growing need for domain-specific LLMs and ML tooling tailored to HPC operations—opportunity for vendors and service providers offering pre-trained, fine-tuned models and MLOps for HPC.
  • Standardization & benchmarks: creation of canonical benchmarks and shared data formats would lower transaction costs, make comparisons meaningful, and accelerate adoption—worth public or consortium funding.
  • Public-good funding: governments and research funders can accelerate development by financing benchmarks, FAIR data initiatives, and reproducibility efforts that individual centers would not internalize.
  • Regulatory / sustainability incentives: energy and emissions reporting or carbon pricing increase the value of energy-saving AI tools, shifting economic balance toward adoption.

• Recommendations for economists, operators, and funders

  • Conduct rigorous ROI studies: quantify TCO (capital + operational + model maintenance) vs. benefits (reduced energy, improved throughput, labor savings).
  • Include energy and carbon metrics in cost–benefit analyses for AI adoption in HPC.
  • Prioritize investments in MLOps, standardized APIs, and benchmark datasets that reduce deployment friction and maintenance cost.
  • Support development and evaluation of hybrid approaches as near-term pragmatic solutions that balance benefits and operational risk.
  • Incentivize domain-specific LMs tailored to HPC tasks (automation, scripting, diagnostics) but require rigorous evaluation for correctness and safety.

Summary takeaway: AI can deliver measurable operational and sustainability value for HPC centers, but economic gains depend on careful accounting for deployment and maintenance costs, hybrid adoption paths, and institutional support for benchmarking and MLOps to lower adoption frictions.