Hot take: if 70% of your work is fine-tuning/inference, unified memory...