GEMM-ArchProfiler: A simulation framework for hardware-level profiling and performance analysis of General Matrix Multiplication in real CNN workloads on heterogeneous CPU architectures

In this paper, the authors present GEMM-ArchProfiler, a simulation framework for evaluating General Matrix Multiplication performance in convolutional neural networks. Targeted at resource-constrained edge and IoT systems, which rely on CPU-based architectures, the framework addresses hardware limitations through optimized workload profiling. Powered by the gem5 simulator, GEMM-ArchProfiler provides insights into memory usage, cache behavior, execution latency, and energy consumption. It integrates customized Darknet libraries to simulate realistic CNN workloads and includes a user-friendly CPU configuration mechanism and event analysis script. This tool bridges workload analysis and deployment, aiding efficient AI implementation on diverse CPU architectures.