Did you run the model on 8 GPUs or more? I cannot get a better result with more GPU, in fact it is worse.