RAPTOR: Practical Numerical Profiling of Scientific Applications

2025-07-07

AI Summary

Modern HPC architectures increasingly feature low-precision hardware (e.g., FP16/INT8) while degrading FP64 capabilities, posing a fundamental accuracy–performance trade-off in scientific computing: naïve precision reduction often triggers numerical failure. To address this, we propose the first transparent numerical profiling framework tailored for multiphysics applications. Built upon LLVM, it enables dynamic instrumentation and simulation of arbitrary precision—including user-defined types—without source-code modification, precisely identifying code regions amenable to safe precision reduction. Our method supports fine-grained error propagation analysis and runtime precision assessment under realistic workloads. Evaluated on four production Flash-X multiphysics applications, it achieves significant computational speedups while preserving numerical reliability. This work bridges a critical gap in the toolchain for precision optimization of complex scientific codes.

Technology Category

• Machine Learning: Hardware-aware ML
• Search and Optimization: Sampling/Simulation-based Search
• Computer Vision: Low Level & Physics-based Vision

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Web performance, measurement, and characterization
• Security and Privacy: Data transparency and provenance
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

The proliferation of low-precision units in modern high-performance architectures increasingly burdens domain scientists. Historically, the choice in HPC was easy: can we get away with 32 bit floating-point operations and lower bandwidth requirements, or is FP64 necessary? Driven by Artificial Intelligence, vendors introduced novel low-precision units for vector and tensor operations, and FP64 capabilities stagnate or are reduced. This is forcing scientists to re-evaluate their codes, but a trivial search-and-replace approach to go from FP64 to FP16 will not suffice. We introduce RAPTOR: a numerical profiling tool to guide scientists in their search for code regions where precision lowering is feasible. Using LLVM, we transparently replace high-precision computations using low-precision units, or emulate a user-defined precision. RAPTOR is a novel, feature-rich approach -- with focus on ease of use -- to change, profile, and reason about numerical requirements and instabilities, which we demonstrate with four real-world multi-physics Flash-X applications.

Problem

• Helps scientists identify code regions for precision lowering
• Profiles numerical requirements and instabilities in applications
• Guides transition from FP64 to lower precision units

Innovation

• Numerical profiling tool for precision analysis
• LLVM-based transparent precision replacement
• User-defined precision emulation for HPC