Ascend to Science: Exploration of AI Chips for Scientific Computing

2026-07-22

AI Summary

This work addresses the challenges of numerical robustness, irregular memory access, and scalability faced by AI accelerators in scientific computing. Leveraging the Ascend 910 NPU as a target platform, the authors propose a co-optimization framework that distinguishes general optimization principles from hardware-specific implementations. The framework integrates heterogeneous execution, mixed-precision numerical methods, precision emulation, hierarchical memory orchestration, and communication-computation overlap to jointly optimize numerical formats, scheduling strategies, and data movement. Evaluated on five representative scientific workloads—including HPL-MxP and LRSVD—the approach demonstrates high performance, strong scalability, and numerical stability, thereby validating its effectiveness and competitiveness for scientific applications on AI accelerators.

Abstract

The rapid rise of AI-oriented accelerators has reshaped compute systems around low-precision tensor engines, raising a practical question for the HPC community: under what conditions can such hardware support scientific workloads that demand numerical robustness, irregular memory access, and scalability? Using the Ascend 910 NPU series as a representative tensor-centric platform, we characterize precision, execution, and memory-hierarchy bottlenecks that hinder the direct deployment of scientific codes. We then develop and evaluate workload-specific mappings across five application studies -- HPL-MxP, LRSVD, SGEMM-cube, PQSim, and SMC-X -- combining heterogeneous execution, mixed-precision numerical formulations, precision emulation, hierarchical memory orchestration, and communication--computation overlap. These studies show that AI-native NPUs can achieve numerical robustness, competitive performance, and satisfactory scalability when numerical formulation, execution placement, and data movement are addressed in a coordinated manner. Our results provide a state-of-the-practice case study of how scientific workloads can be adapted to tensor-centric architectures, while distinguishing transferable optimization principles from Ascend-specific implementation details.

Problem

• AI accelerators
• scientific computing
• numerical robustness
• memory access
• scalability

Innovation

• tensor-centric architecture
• mixed-precision computing
• hierarchical memory orchestration
• numerical robustness
• heterogeneous execution