GPU workloads are very parallel. By throwing more transistors at the problem while lowering clock rates you can get pretty good performance even in a constrained power budget.