GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model

Meta has optimized the training process for its Generative Ads Recommendation Model, the foundation model powering advertisement targeting on Facebook and Instagram, to double its end-to-end efficiency. By scaling the model to run on thousands of high-end GPUs, the company has reduced the computational resources required for its recommendation infrastructure. This advancement allows Meta to iterate on its ad delivery systems more rapidly and maintain performance as the complexity of its underlying machine learning models continues to grow.