Skip to content
Science

December 2, 2025 (9.05 pm. ET for Asia-USA)

December 2, 2025 (9.05 pm. ET for Asia-USA)

Present

  • Dora Cheng, Gary Mazzaferro, Geoffrey Fox, Gregor von Laszewski, Hao Lu, Michael A., Piotr Luszczek, Satoshi Iwata, Victor Lu

Google Meet Notes

Benchmark Paper Update

  • The primary benchmark paper is "in principle finished" and has been refined using feedback from a Claude-generated report.
  • It is currently awaiting formal approval from Fermilab, which is required before publication on arXiv to ensure compliance with export control rules.
  • Gary Mazzaferro and new member Hao Lu offered to conduct a final review of the manuscript, with Gary aiming to complete his review by Monday.
  • The paper establishes a formal definition (a superset) for benchmarking that captures complex mixed AI and High-Performance Computing (HPC) tasks.
  • The group agreed that the current paper is the final version and any further work will be a follow-up paper, such as one focusing on database aspects led by Victor.
    Time Series and LLMs Discussion
  • The discussion centered on time series systems like Salesforce's Merlion (which includes models like Morai and ETS Former), TimeGPT (LLM-based, part of the Neural Forecast portal), and Amazon's Kronos (LLM architecture trained on numerical data).
  • Geoffrey Fox and Gregor von Laszewski agreed that LLMs generally do not predict time series well because they learn language structure rather than time structure.
  • Gary Mazzaferro is looking at time series neural networks for a cholera prediction application for South Saharan Africa, noting the complexity of combining time series with geospatial factors like waterway flows, animal migration, and weather conditions.
    New Members and Introductions
  • Geoffrey Fox welcomed three new members: Hao Lu from Salesforce (Responsible AI team), Dora Cheng from Super Micro, and Michael from the state government of New York.
    SNIA Podcast and Hardware Issues
  • Gary Mazzaferro is seeking a second panel guest for a SNIA web podcast on data management for HPC and AI applications.
  • Geoffrey Fox offered to forward the contact information of his "best student" who is working on data movement for LLMs in both training and inference.
  • Gary Mazzaferro highlighted a current industry issue: low GPU memory utilization (around 60%) in current hardware architectures and the need to motivate the PCI SIGs to add cache coherency signals to the PCI bus to address this.
    Other Benchmarking Efforts
  • Gregor von Laszewski reported on a practical performance benchmark on YOLO/DarkNet with a Florida group, using the group's carpentry paper to formalize the benchmark and creating an MLPerf-compatible logging infrastructure. He plans to present on this work in the second week of January.
  • The next MLC Science WG meeting is scheduled for next Wednesday at 11:05 Eastern.

New Members

Discussion