sow_GMU_2025rv2.pdf

This Statement of Work (SOW) outlines requirements for GPU Modeling Research of Smart Modeling and Simulation for HPC (SMASH) at Brookhaven National Laboratory, with performance duration from January 1, 2025 to September 30, 2025. The project aims to develop machine learning (ML)-based performance models for GPUs used in scientific computing and ML workloads, extending the applicability of ML-based modeling approaches currently under development at Brookhaven National Laboratory.

The contractor must complete three main tasks: 1) Design an ML technology to model parallel GPU execution with reasonable accuracy and reduced computation overhead, 2) Improve ML-based GPU model accuracy through neural architecture search, and 3) Demonstrate use cases for the proposed ML-based GPU models in design space exploration. Deliverables include a codebase implementing the proposed ML-based GPU models due 200 days after receipt of order (ARO) and a technical paper describing the implementation and evaluation due 220 days ARO. The schedule outlines specific milestones including model development for single thread warp execution (2 months), parallel warp and thread block execution (3 months), model optimization (2 months), use case exploration (2 months), and technical paper writing (2 months).