Skip to content
Science

May 18, 2022

May 18, 2022

Present

Junqi Yin, Gregor von Laszewski, Juri Papay, Gregg Barrett, Farzana Yasmin Ahmad, Aristeidis Tsaris, Piotr Luszczek, Cade Brown, Geoffrey Fox, Tony Hey, Arjun Shankar, Bala Desinghu, Rushil Anirudh, Hyojin Kim

Tentative Agenda

New member introductions

  • We were joined by two Lawrence Livermore researchers Rushil Anirudh and Hyojin Kim who each described an open Lawrence Livermore benchmark dataset and given in next item.

Livermore Presentations

  • Rushil Anirudh described a Fusion benchmark LLNL_ICF_Readme.pdf. There are 10000 points in the open dataset but internally there are 100 million such data points.
  • This surrogate benchmark predicts images from input plasma parameters but most interesting is the inverse problem (images predict plasma condition)
  • One problem is creating synthetic data and measure diversity
  • One question is what is the best metric
  • One use is design optimization
  • Here the issue is not scaling but rather problem complexity
  • Hyojin Kim described a computed tomography material science benchmark LLNL_D4DCT_MLCommons_Updated.docx.
  • The tomography data needs to be analyzed to reconstruct the deformed image
  • One dataset is 80x80x80 3D images with 10 frames (time instances) each. One has the Xray measurement plus the ground truth deformed images. One needs to improve the reconstruction.
  • A further dataset is 256x256x256 with 60 frames each.
  • Peak signal-to-noise ratio (PSNR) measures quality of reconstruction
  • The large images stress memory on GPUs and as well as improving PSNR one needs to scale solution to run on multiple GPUs.
  • Tom Gibbs NVIDIA later made interesting comments that extend this area

Our Paper and Policy Documents

  • Arjun made an interesting point for image based benchmarks. Here the deep learning technology like RESNET are well established and a key step is to feed in domain knowledge (physics informed networks).
  • Juri and Gregor described the paper and policy document. Both are essentially finished
  • Current work in both documents is in section that gives details of the 4 benchmarks.
  • The benchmarks are currently in https://github.com/laszewsk/mlcommons but need to be moved into the MLCommons GitHub
  • We need to give tgo Peter Mattson and David Kanter to review our work
  • We discussed publication options; one is an ISC workshop http://www.icl.utk.edu/\~luszczek/conf/2022/h3/ . Luszczek and Hey will investigate

Any Other Business