Day 1: Jack Dongarra - High Performance Computing and Responsibly Reckless Algorithms

Day 1: Jack Dongarra - High Performance Computing and Responsibly Reckless Algorithms

by Jack Dongarra

Formal & Physical Sciences Computing & Cybersecurity UYFComputer architecture and logic designUYFPParallel processing
🎙 Jack Dongarra 👥 824 📅 November 5, 2025 ⏱ 32 min 👁 201 📄 expert opinion 🧭 2026-08-16
Available in: English (current) Français

Keywords

HPCmixed precisiontensor coresOzaki schemeexascale

Summary

Jack Dongarra, a leading figure in high-performance computing, presents a talk on the evolution of HPC and introduces the concept of ‘responsibly reckless’ algorithms. He traces the history from vector machines to current GPU-accelerated systems, highlighting the shift towards commodity processors and the dominance of NVIDIA in accelerators. He discusses the growing gap between computation speed and data movement, and the trend of cloud providers designing custom hardware. He presents the Top500 list, noting the US’s El Capitan as the fastest, and the absence of China due to sanctions. He then focuses on mixed-precision arithmetic, showing how lower precision can dramatically speed up computations, and introduces the Ozaki scheme for using integer tensor cores to emulate floating-point operations. He emphasizes the need for algorithms that are fast but may occasionally fail, requiring checks and fallbacks. He concludes by noting that hardware is constantly changing and algorithms must adapt.

148 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the current state and future directions of HPC, with concrete examples and data. Dongarra’s argumentation is solid, based on his extensive experience and recent market data. He effectively explains complex concepts like mixed precision and the Ozaki scheme, making a compelling case for their adoption. However, some claims, such as the future of quantum computing, are presented as personal opinions rather than evidence-based conclusions.

Scientific Rigor, Source Quality, Title Accuracy

Dongarra is a highly credible source, and the talk references real systems and benchmarks. The sources cited are primarily the Top500 list and his own research, which are reliable. The title accurately reflects the content, and the talk is well-structured. No public comments were provided for analysis.

132 words

Title / Content Match

The title accurately reflects the content, focusing on HPC evolution and the concept of responsibly reckless algorithms.

Quality & Reliability

9/10

Presentation by a leading expert in HPC, with concrete data and references to ongoing research, though some claims are based on personal perspective and market data.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

  • Top500 List — Official list of the world's fastest supercomputers, referenced in the talk.

Contribution & Novelties

The talk introduces the concept of ‘responsibly reckless’ algorithms, which combine fast but potentially unstable methods with checks and fallbacks. It also highlights the Ozaki scheme, a novel approach to using integer tensor cores for floating-point operations, and discusses the trend towards mixed-precision computing. This provides a forward-looking perspective on HPC.

Pour aller plus loin :

  • Mixed-precision arithmetic — Overview of mixed-precision techniques.
  • Tensor core — Explanation of tensor cores in GPUs.
  • Ozaki scheme — Original paper on the Ozaki scheme for matrix multiplication using integer arithmetic.

87 words

Radar Profile

The radar profile shows high scores in information quality, technical level, and reliability, with a slightly lower score in information quantity due to the talk's focused scope. This indicates a highly informative and credible presentation, though not exhaustive in covering all aspects of HPC.

Reliability 9/10