Accel-SIM 2.0: Validated GPU Simulation with Full Hopper Support (github.com)

Accel-SIM 2.0 has been announced, introducing extensive support for NVIDIA's Hopper architecture, enabling enhanced GPU simulations that are crucial for modern AI workloads. This validated framework allows for cycle-level performance modeling by tracing actual SASS executions from NVIDIA hardware, achieving an impressive accuracy of 99% Pearson correlation against real H100 silicon in its validation across over 34,000 kernel instances. The update facilitates seamless GPU simulation with features such as support for Tensor Memory Accelerator (TMA), asynchronous Warp Group MMA (WGMMA), and advanced synchronization techniques, making it a powerful tool for researchers and engineers in the AI/ML community. The significance of Accel-SIM 2.0 lies in its compatibility with popular machine learning frameworks like PyTorch and vLLM, allowing users to trace and simulate large language models with minimal effort. Users can simply pull models from Hugging Face and run simulations with a few lines of Python code, eliminating the need for manual kernel adaptations. This robust capability addresses the growing complexity of modern AI applications and streamlines the research and development process, paving the way for more efficient GPU utilization and innovation in AI model architectures.