Research5h ago
Meta Doubles Efficiency of Ads Foundation Model
A 20–25% Model FLOPs Utilization rate marks the new efficiency peak for Meta's Generative Ads Recommendation Model. Engineers achieved this by co-designing kernels, precision, and networking to scale training FLOPs 4x in one year. The GEM architecture blends recommendation-domain data with LLM scale. This optimization proves that specialized infrastructure can drastically reduce training costs for hybrid models.