MT4G: A Tool for Auto-Discovery of NVIDIA and AMD GPU Compute, Memory Topologies (arxiv.org)

Researchers released MT4G, an open-source, vendor-agnostic tool that automatically discovers GPU compute and memory topologies for NVIDIA and AMD devices—filling a gap analogous to hwloc for CPUs. MT4G combines existing vendor APIs with a library of 50+ microbenchmarks and statistical methods (notably Kolmogorov–Smirnov testing) to infer otherwise unavailable hardware attributes such as cache sizes, bandwidths, and physical interconnect/layout. The approach is designed to be robust across vendor idiosyncrasies: benchmarks probe latency/bandwidth patterns and the statistics detect consistent structural signatures rather than relying on brittle heuristics or proprietary queries. The paper demonstrates MT4G across ten GPUs and integrates it into three practical workflows—GPU performance modeling, GPUscout bottleneck analysis, and dynamic resource partitioning—showing concrete gains in modeling accuracy, automated bottleneck identification, and portable resource management. For the AI/ML and HPC communities, MT4G promises more reliable hardware-aware scheduling, tuning, and simulation across heterogeneous clusters, enabling automated optimization pipelines that no longer depend on vendor-specific tooling or manual reverse engineering of memory/compute layouts. Code, data, and demos accompany the release for adoption and reproducibility.