Repository logo
Log In(current)
  1. Home
  2. Colleges & Schools
  3. Graduate School
  4. Masters Theses
  5. Design and Evaluation of a Standardized Interface for AMD GPU Performance Metrics
Details

Design and Evaluation of a Standardized Interface for AMD GPU Performance Metrics

Date Issued
May 1, 2026
Author(s)
Woun, Dong Jun
The University of Tennessee
Advisor(s)
Catherine Schuman
Additional Advisor(s)
Heike Jagode
James Plank
Permanent URI
https://trace.tennessee.edu/handle/20.500.14382/57857
Abstract
Modern high-performance computing (HPC) systems increasingly rely on graphics processing unit(GPU) accelerators, making accurate and low-overhead monitoring of hardware behavior essential for performance analysis, energy optimization, and reliability studies. Differences and updates in vendor application programming interfaces(API) make cross-platform measurement difficult and force researchers to maintain vendor-specific code. The Performance API (PAPI) provides a common interface for accessing hardware performance metrics on central processing units(CPU), GPUs and interconnects. However, AMD's transition from ROCM System Management Interface (ROCm SMI) to the AMD System Management Interface (AMD SMI)---deprecating the former interface previously used by PAPI to access GPU metrics---underscores the need for a robust and vendor-neutral interface for AMD GPUs and accelerated processing units (APU). This thesis presents the design and evaluation of a portable performance monitoring interface for AMD GPUs that abstracts vendor-specific API differences while maintaining low-overhead. The proposed interface ensures measurement stability across software and hardware versions. The novelty lies in a probe-based, automatic discovery mechanism that ensures access to available metrics and consistency across GPU generations. The implementation is built on AMD SMI and integrated into the PAPI component model. The new component automatically detects which metrics are supported by a specific device and driver and exposes only those metrics. The interface provides access to metrics for monitoring power and energy, temperature, engine activity, and interconnect bandwidth/link state. Experimental results confirm that the proposed interface maintains near-native measurement overhead and produces consistent results across different AMD GPU models and ROCm versions, validating its portability and reliability. In summary, this work enables portable, low-overhead access to AMD GPU/APU performance metrics, reducing the maintenance burden of vendor-specific instrumentation, and improving reproducibility in performance studies. It further enhances the portability and reliability of GPU/APU performance monitoring, providing a sustainable foundation for cross-platform performance analysis in heterogeneous HPC environments.
Subjects

Hardware Systems

Computer and Systems ...

Computer Science

High-performance comp...

Performance API (PAPI...

AMD System Management...

GPU performance metri...

hardware performance ...

PCIe/XGMI link metric...

RAS/ECC reliability m...

Degree
Master of Science
Major
Computer Science
File(s)
Thumbnail Image
Name

WOUN-PRIMARY-2026.pdf

Size

2.5 MB

Format

Adobe PDF

Checksum (MD5)

2eca08b7380369269a8027b95b99620c


University Libraries

1015 Volunteer Boulevard
Knoxville, TN 37996
865-974-4351

Map & Directions
Donate to the Libraries
  • About
  • John C. Hodges Society
  • Speaking Volumes magazine
  • Outreach
  • Directory
  • Employment
  • Policies
  • Library Intranet
University of Tennessee power T logo

The University of Tennessee, Knoxville
Knoxville, Tennessee 37996
865-974-1000

Events
A-Z
Apply
Privacy
Map
Directory
Give to UT
Accessibility

Built with DSpace-CRIS software - Extension maintained and optimized by 4Science