Deucalion: Supercomputing for Generative AI Development (use AI to test AI)

Supercomputer Deucalion accelerates generative AI research, enabling rigorous statistical and probabilistic prompt testing to ensure system quality

Deucalion, Portugal's 10-petaflop “green” supercomputer, offers unparalleled computational power and efficiency, which is ideally suited to the demanding workloads of generative AI. Co‑funded by the EuroHPC Joint Undertaking and Portugal’s Foundation for Science and Technology (FCT), Deucalion provides national and European researchers with access to a versatile high‑performance computing (HPC) platform optimised for everything from large‑scale model training to exhaustive prompt quality assurance. By harnessing its ARM, x86, and GPU partitions, AI practitioners can perform Monte Carlo-style prompt sampling, probabilistic inference evaluations, and energy-efficient large language model (LLM) experiments well before deployment, thereby reducing the risks of bias, hallucinations, and system failures in production.

Generative AI and the Need for HPC

Generative AI models create new content by identifying and sampling from statistical patterns in massive training datasets. These foundation models underpin applications ranging from automated text generation to image synthesis, yet their outputs can vary unpredictably due to inherent probabilistic mechanisms. Ensuring that a generative system behaves reliably across diverse prompts, therefore, requires extensive computational experimentation, far beyond the capabilities of conventional servers. High-performance computing infrastructures, designed for parallel processing and large memory bandwidth, become essential for managing the combinatorial explosion of prompt variations and performance metrics.

Image: Portuguese supercomputer Deucalion joins TOP 500 of the world's most efficient supercomputers, FCT

Deucalion: Portugal’s National and European HPC Asset

Located at the Minho Advanced Computing Centre in Guimarães, Deucalion was inaugurated in September 2023 and is co‑sponsored by FCT and EuroHPC. It integrates:

• ARM Partition: 1,632 Fujitsu PRIMEHPC FX700 nodes (A64FX processors), 16 GB HBM each, peaking at 5 PFLOPS.

• x86 Partition: 500 Atos Bull Sequana X440 nodes (AMD EPYC Rome 7742), 256 GB RAM each, peaking at 2.3 PFLOPS.

• GPU Partition: 33 Sequana E410 nodes with four Nvidia A100 GPUs each, 80 GB HBM per GPU, peaking at 2.7 PFLOPS.

  Altogether, Deucalion achieves a peak of 10 PFLOPS and features in the TOP500 (ranked #219) and Green500 (#80 for energy efficiency) lists, showcasing both raw power and sustainability.

Statistical and Probabilistic Prompt Testing

Generative AI relies on sampling probability distributions to produce outputs, making statistical and probabilistic testing indispensable. By executing large batches of slightly varied prompts across multiple model instances, researchers can:

• Estimate Response Distributions: Quantify variability and confidence intervals for outputs under different temperature and top‑k settings (testpublishers.org).

• Detect Bias and Hallucinations: Use statistical anomaly detection to flag outputs that deviate significantly from expected patterns.

• Optimise Prompt Templates: Employ Bayesian optimisation or particle‑based Monte Carlo methods to identify prompt structures that maximise quality and relevance.

Deucalion’s massive parallelism enables these experiments to run in hours rather than days or weeks, thereby accelerating the prompt engineering cycle.

Image: How is Generative AI different from Traditional tech, American Academy of Actuaries, by Venkat Seshadri and Darin Hornsby

Ensuring System Quality Before Production

Advanced workflows, such as the proposed “Digital Human Tester” (DHT), emulate human testers by automatically generating, executing, and validating test cases for Generative AI systems, as described by Actuary.org. When deployed on Deucalion, DHT agents can orchestrate thousands of test scenarios concurrently, providing comprehensive coverage of edge cases and performance bottlenecks. This pre‑production validation pipeline minimises risks in customer‑facing applications—especially critical in sectors like finance, healthcare and education where errors can have significant consequences.

By combining Deucalion’s scalable HPC capabilities with rigorous statistical and probabilistic prompt‑testing methodologies, organisations can certify generative AI systems for reliability, fairness and safety before they reach production. As generative AI continues to permeate every industry, the symbiosis between supercomputing and AI research, embodied by Deucalion, will be pivotal in transforming innovation into trustworthy, real-world solutions.