GEM TRAINING: HOW META DOUBLED THE EFFICIENCY OF ITS LLM-SCALE ADS FOUNDATION MODEL

Summary from the TLDR Data newsletter

Meta scaled its ads recommendation foundation model 4x in 12 months and doubled training efficiency to 20–25% MFU through five techniques: jagged flash attention to eliminate padding waste, a unified attention kernel with 2–3.5x speedup, MXFP8 precision training, SM-free collective communication that reclaims ~23 SMs, and a load-balancing shuffle that adds 4% efficiency at zero communication cost.