Centimani - Enabling Fast AI Accelerator Selection for DNN Training

Explore a conference talk from USENIX ATC '24 that introduces Centimani, a novel performance predictor designed to streamline AI accelerator selection for DNN training. Delve into the challenges of choosing optimal hardware for deep neural network models as AI-specific accelerators proliferate. Learn how Centimani accurately predicts DNN training throughput across various accelerators, enabling informed decision-making based on objectives like performance or cost-efficiency. Discover the innovative approach involving memory estimation and decoupled performance models to determine ideal batch sizes and forecast execution times. Examine the validation results showcasing Centimani's high prediction accuracy for both single and multi-device training scenarios across multiple DNN models and accelerators.