“On GPUs, Model Flops Utilization (MFU) is typically between 20% and 50% for real workloads.”