cost-effectiveness of GPU clusters