4th Data Prefetching Championship (DPC4) - HPCA 2026 - Sydney - Feb 1, 2026

4th Data Prefetching Championship (DPC4) - HPCA 2026 - Sydney - Feb 1, 2026

🎙 Onur Mutlu Lectures 👥 64K 📅 February 1, 2026 ⏱ 255 min 👁 2K 📄 documentary 🧭 2026-08-15
Available in: English (current) Français

Keywords

prefetchingcachecomputer architectureDPC4HPCA

Summary

This video is a recording of the 4th Data Prefetching Championship (DPC4) workshop held at HPCA 2026 in Sydney. The workshop begins with an introduction by Rahul Bera, one of the organizers, who explains the motivation for the championship, highlighting the continued importance of prefetching in modern processors, as evidenced by the 5.3% performance improvement attributed to hardware prefetchers in ARM’s Neoverse V2. He outlines the goals of DPC4, which include evaluating innovative prefetching ideas under a common framework using the ChampSim simulator, with a focus on new workloads such as AI/ML inference and graph processing. The evaluation involved 610 single-core and 484 four-core workloads, consuming over 50,000 CPU core hours. The results show a 5.2% performance improvement for the winning submission over a baseline that already includes L1/L2 prefetchers, with notable gains in AI/ML workloads. The workshop then features an invited talk by Leor Peled from Huawei, who discusses the history and future of prefetching, emphasizing the challenges of irregular patterns and the potential of semantic prefetching. The video also includes presentations of the eight competing prefetching proposals and concludes with the announcement of the winners. The recording is technical and aimed at an audience familiar with computer architecture.

200 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides valuable insights into the current state of data prefetching research. The introduction by Rahul Bera effectively argues for the continued relevance of prefetching by citing concrete performance improvements in commercial processors. The presentation of the championship’s methodology and results offers a comprehensive overview of the evaluation process and the performance gains achieved. Leor Peled’s invited talk is particularly valuable, as it provides a historical perspective on prefetching and challenges the audience to think about future directions, such as semantic prefetching and handling irregular patterns. The argumentation is solid, grounded in data and experience, and encourages critical thinking about the field’s evolution. However, the video is a workshop recording, so the presentations are not peer-reviewed in the traditional sense, but they are based on rigorous evaluation and expert knowledge.

Scientific Rigor, Source Quality, Title Accuracy

The video demonstrates scientific rigor through the structured evaluation framework of DPC4, which uses a common simulator (ChampSim) and a large set of workloads. The organizers emphasize the public release of all source code, papers, and results to ensure reproducibility. The sources cited include the DPC4 website and the main program page, which provide access to the workshop details and presumably the papers. The title accurately reflects the content, as the video is indeed a recording of the DPC4 workshop. The invited talk by Leor Peled references his own work and general knowledge of the field, but does not cite specific external sources. Overall, the video maintains a high standard of scientific rigor, with clear methodology and transparency.

264 words

Title / Content Match

The title accurately reflects the content: a recording of the 4th Data Prefetching Championship workshop at HPCA 2026.

Quality & Reliability

8/10

The video is a recording of a scientific workshop (DPC4) organized by established researchers, featuring peer-reviewed submissions and industry talks. The content is technical and presented by experts. However, it is a workshop recording, not a peer-reviewed publication itself, and the evaluation results are presented by the organizers, which may introduce bias.

Key Moments

Cited Sources

  • DPC4 Workshop Website — Official website for the 4th Data Prefetching Championship, providing details on the workshop, rules, and results.
  • DPC4 Main Program — Main program page for DPC4, listing the schedule and presentations.

Concurring Sources

  • ChampSim Simulator — The simulation framework used for DPC4 evaluations, widely used in computer architecture research.
  • ARM Neoverse V2 Presentation — Reference to the ARM Neoverse V2 processor, which the introduction cites as showing significant performance gains from prefetching.

Contribution & Novelties

The video provides a comprehensive overview of the 4th Data Prefetching Championship, showcasing the latest research in data prefetching. It introduces new workloads representative of modern AI/ML and graph processing, and presents a rigorous evaluation framework. The invited talk by Leor Peled offers a unique perspective on the future of prefetching, emphasizing the need for semantic approaches. The championship itself contributes to the field by providing a common platform for evaluating and comparing prefetching techniques, and by releasing all code and results for further research.

Pour aller plus loin :

  • Data Prefetching Championship — Official website of the previous championship, providing context and historical results.
  • ChampSim Simulator — The simulation framework used in DPC4, open-source and widely used in academia.
  • Semantic Prefetching — Wikipedia article on prefetching, providing general background and related concepts.
  • Machine Learning for Prefetching — A survey paper on ML-based prefetching techniques, relevant to the discussion of neural network prefetchers.

153 words

Radar Profile

The radar profile shows high scores in technical level and information quality, reflecting the advanced and detailed content. The quantity of information is also high, given the extensive coverage of the workshop. The global reliability is strong, but slightly lower due to the nature of the content as a workshop recording rather than a peer-reviewed publication.

Reliability 8/10