Meta Doubles Ads Model Training Efficiency at LLM Scale
Meta's Generative Ads Recommendation Model (GEM) now trains on thousands of latest GPUs, achieving 20‑25% model FLOPs utilization and a 4× increase in training FLOPs, effectively doubling end‑to‑end efficiency. They did this via custom kernels, ultra‑low‑precision training, and a 5‑dimensional topology‑aware parallelism stack.