Quail is an execution engine for AI-SQL that extends SQL with LLM function calls. The article explains how to estimate latency of AI-powered filter operations using the roofline model, which computes arithmetic and memory requirements then divides by GPU hardware limits to obtain speed-of-light estimates for query execution.
This article analyzes the performance scaling of Jacobi and Gauss-Seidel stencil kernels implemented in Fortran using roofline analysis. The authors observe that parallelization scaling differs significantly between smaller (512×512) and larger (2048×2048) grids, prompting a simplified hardware model based on compute capacity and memory bandwidth to explain the performance bottlenecks.