MT4G: A Tool for Reliable Auto-Discovery of NVIDIA and AMD GPU Compute and Memory Topologies

2025-11-08

Proceedings of the SC '25 Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis

AI Summary

GPU topology information—such as compute/memory hierarchy, cache sizes, interconnect bandwidths, and physical layout—is severely fragmented, incomplete, and vendor-locked, hindering performance modeling and optimization in HPC and AI systems. To address this, we propose the first cross-vendor (NVIDIA/AMD), open-source framework for automatic GPU hardware topology discovery. Our approach integrates CUDA/HIP APIs with over 50 customized microbenchmarks and employs statistical validation—including Kolmogorov–Smirnov tests—to infer non-programmable hardware attributes with high confidence. We validate the framework across ten mainstream GPUs, demonstrating broad compatibility and measurement accuracy. Furthermore, we integrate it into three critical workflows: analytical performance modeling, bottleneck analysis, and dynamic resource partitioning. This integration significantly enhances system-level hardware awareness and improves resource utilization efficiency.

Technology Category

• Machine Learning: Hardware-aware ML
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization
• Search and Optimization: Distributed Search

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Web performance, measurement, and characterization
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Web Mining and Content Analysis: Topic discovery and tracking

Abstract

Understanding GPU topology is essential for performance-related tasks in HPC or AI. Yet, unlike for CPUs with tools like hwloc, GPU information is hard to come by, incomplete, and vendor-specific. In this work, we address this gap and present MT4G, an open-source and vendor-agnostic tool that automatically discovers GPU compute and memory topologies and configurations, including cache sizes, bandwidths, and physical layouts. MT4G combines existing APIs with a suite of over 50 microbenchmarks, applying statistical methods, such as the Kolmogorov-Smirnov test, to automatically and reliably identify otherwise programmatically unavailable topological attributes. We showcase MT4G's universality on ten different GPUs and demonstrate its impact through integration into three workflows: GPU performance modeling, GPUscout bottleneck analysis, and dynamic resource partitioning. These scenarios highlight MT4G's role in understanding system performance and characteristics across NVIDIA and AMD GPUs, providing an automated, portable solution for modern HPC and AI systems.

Problem

• Automating GPU topology discovery for HPC and AI systems
• Addressing vendor-specific limitations in GPU information access
• Providing reliable identification of unavailable GPU topological attributes

Innovation

• Open-source tool auto-discovers GPU topologies
• Combines APIs with microbenchmarks for reliability
• Applies statistical methods to identify hidden attributes