Shadowfetch News — AI news. Real coverage.

Infrastructure

Meta says GEM doubled efficiency while scaling its ads foundation model at LLM scale

Meta published its training notes for GEM, outlining how infrastructure and kernel changes delivered a notable jump in training efficiency while scaling ads recommendation capacity across thousands of GPUs.

a rack of servers in a server room
Photo by Kevin Ache on Unsplash

Meta's Engineering at Meta post on August 3, 2026 says GEM, the foundation model behind Instagram and Facebook ads recommendations, now runs at LLM scale across several thousand latest-generation GPUs.

The post says Meta co-designed kernels, precision, parallelism, networking, and memory to double GEM's end-to-end efficiency to 20–25% Model FLOPs Utilization (MFU).

It also says training FLOPs scaled up 4x in 12 months under that same efficiency-focused effort.

The same writeup describes GEM as a hybrid system with trillions of sparse embedding parameters and billions of dense parameters, with recommendation data patterns like jagged user-history sequences making standard LLM training assumptions inefficient.

To address this, Meta says it built custom kernels such as Jagged Flash Attention, Generalized Dot-Product Attention, and BlockAttention, and added low-precision MXFP8 attention and MLP paths.

For distributed scale-out, Meta cites 5D parallel strategies across dense and sparse dimensions and topology-aware networking, including a multi-tier cluster view, to reduce communication bottlenecks while tracking both compute and scaling efficiency together.

Sources

  1. GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model
  2. Meta’s Generative Ads Model (GEM): The Central Brain Accelerating Ads Recommendation AI Innovation
  3. Collective Communication for 100k+ GPUs

From Shadowfetch