Lambda Labs Pushes Model MFU Past 60% on Blackwell
September 14, 2026
Lambda engineers achieved over 60% Model FLOPS Utilization (MFU) for Llama 3.1 models on NVIDIA Blackwell GPUs. The improvement was realized through new parallelism strategies and memory overhead management.
HOW THIS AFFECTS YOU
●
builderThis sets a new baseline for training and inference efficiency on next-gen hardware.
●
researcherYou can leverage these parallelism strategies to scale larger training runs more effectively.