Skip to content
Science

November 30, 2022

November 30, 2022

Present

Tony Hey, Wesley Brewer, Geoffrey Fox, Piotr Luszczek, Juri Papay, Christine Kirkpatrick, Farzana Yasmin Ahmad, Aristeidis Tsaris, Murali Emani, Tom Gibbs

Apologies

Jeyan Thiyagalingam, Gregor von Laszewski, Gregg Barrett,

Tentative Agenda

  • Any new members - None
  • SC22 BOF Report
  • Futures -- new benchmarks
  • Short presentations on possible benchmarks by Wesley Brewer, and Piotr Luszczek
  • AOB

Discussion

  • SC22 BOF was popular and had plenty of questions
  • Need to follow up on MLCommons Public Relations re announcement of availability of benchmarks
  • Draft is MLCommons Science Blog post draft

Presentation by Piotr Luszczek

  • sciml_autotune_gp_Luszczek.pdf describes a Gaussian process based surrogate aimed at “Autotuning” – choosing best parameters to run a particular code (e.g. BLAS member) on a particular machine.
  • Gaussian processes like an infinitely wide neural network
  • Uses EGO Efficient Global Efficient Global Optimization (EGO) — SMT 1.4.dev documentation for non convex no derivative optimization
  • Can choose Floating point Integer or categorical state for parameters
  • Use Latin Hypercude as Space filling sample Latin hypercube sampling - Wikipedia
  • Hierarchy exploited
  • Around 30 samples used; this number to be studied and also improving performance
  • Looking at one-shot encoding
  • As problems form classes use transfer learning among related problems

Presentation by Wesley Brewer

  • MLCommons-Brewer-12-03-22.pdf describes an Inference network used in Computational Dynamics simulations to look up aspects of model as used in helicopter rotor aerodynamic codes
  • Typical paper is https://ieeexplore.ieee.org/abstract/document/9652868
  • There are three models: an LSTM for simple lookups and two larger models: CNN and TCNN (Temporal Convolutional Neural Network). In latter two, velocities are input at particular times and the pressures are returned.
  • The work SMIBench includes both the surrogate neural networks and a sophisticated Inference server whose performance was optimized to reduce Inference cost to about 10% of the CFD code.
  • Physics Informed Networks were looked at but not used due to difficulties with boundary conditions.