Models & Labs
Meta Doubles Efficiency of GEM Ads Model
Meta AIhigh confidence
Why it matters
- →Meta's advancements in GEM training efficiency set a new standard for large-scale recommendation models.
- →The co-design of hardware and software demonstrates a novel approach to overcoming LLM-scale training challenges.
- →These innovations could lead to more personalized and efficient ad recommendations on major social platforms.
Meta has doubled the training efficiency of its Generative Ads Recommendation Model (GEM), which is crucial for ad recommendations on Instagram and Facebook. By integrating hardware and software innovations, Meta achieved a fourfold increase in training FLOPs over the past year. The improvements include a custom kernel library and ultra-low precision training optimized for the latest GPUs. These advancements address the unique challenges of training recommendation systems at LLM scale, enhancing the model's ability to process complex ad and user interaction data efficiently.
Read original
