Meta GEM Training Efficiency Doubled to 20-25% MFU; Custom Kernels Close Recommendation-LLM Gap

Meta published August 3, 2026 engineering details on how it doubled end-to-end training efficiency of its Generative Ads Recommendation Model (GEM) to 20-25% Model FLOPs Utilization (MFU) while scaling training FLOPs 4x over 12 months. GEM is the largest foundation model for recommendation systems ever built, trained at LLM scale on thousands of GPUs. The model powers ad recommendations across Facebook and Instagram and has delivered 5% conversion lift on Instagram and 3% on Facebook Feed since launch, with Q3 gains doubling relative to Q2.

GEM training presents unique challenges not found in LLM workloads: highly variable sequence lengths (users' activity history ranges from hundreds to tens of thousands of tokens), asymmetric attention patterns (long sequence history, short ad-user interaction windows), and memory-bound sparse operations. Meta built custom GPU kernels—Jagged Flash Attention (JFA), Generalized Dot-Product Attention (GDPA), BlockAttention—that operate directly on jagged tensors and mixed ultra-low precision (MXFP8) training tuned for recommendation workloads. A topology-aware 5D parallelism scheme with SM-free collectives co-designed around Meta's multi-tier network reduced communication overhead.

For infra builders: the result is transferable. Meta's proof that recommendation-scale models can follow LLM-like efficiency scaling laws (if you co-design kernels + precision + parallelism together) applies to any heterogeneous foundation model combining sparse embeddings with dense transformers. The 4x FLOP scaling in 12 months and doubled MFU show that systems-level innovation can unlock efficiency gains comparable to architecture-level breakthroughs—relevant for teams training multi-task or multi-modal models at scale.