PPT-GPU 2.0: A Validated, Accurate, and Scalable Modeling of NVIDIA GPUs

Performance modeling and simulation tools (ModSim) for GPUs continue to be indispensable for performance optimization and a deeper understanding of their complex behaviors and architectures. We detail PPT-GPU 2.0, highlighting significant updates and improvements to PPT-GPU. These include a detailed modeling of the Ampere architecture, including shared memory and tensor cores, while extending coverage across Volta, Turing, and Ampere. Furthermore, we introduce a new GPU occupancy model, “Balance Model”. New performance metrics are also incorporated, including the maximum/minimum number of blocks executed concurrently per streaming Multiprocessor, and the active cycles for each execution pipeline. We evaluate PPT-GPU 2.0 on A100 GPUs, an Ampere architecture. We validate the results using a diverse set of benchmarks compared against real hardware, utilizing NVIDIA Nsight Compute, Accel-Sim (a cycle-accurate simulator), and HyFiSS. PPT-GPU 2.0 demonstrates notable accuracy improvements. For instance, it achieves an average error of 16.4% for active cycles, 9.3% for standard execution pipelines and 5.7% for tensor cores, while maintaining high accuracy for memory hierarchy (8% L1, 21.4% L2 hit rates, and 16.4% for IPC) and execution pipeline modeling. It achieves a 33.51x speedup in simulation time compared with Accel-Sim.