GPU-Tile-Sim: A Tile-Centric GPU Simulation Framework for LLM Hardware-Software Co-Design

GPU-Tile-Sim addresses the problem that existing GPU performance models are either too costly to adapt or too coarse for modern LLM kernels. The method represents kernel execution as a warp-level tile graph, capturing tile operations and dependency constraints, with an automatic frontend and graph-driven backend. Experimental evidence on A100 and H100 across GEMM, attention, and end-to-end LLM inference shows high accuracy (MAPE 1.22%–8.71%), with preliminary validation on Blackwell. This matters because it enables efficient hardware-software co-design for evolving GPU architectures and LLM workloads.