Hmmm... According to the benchmark configuration they used powersave mode. In general you want to use a fixed frequency when benchmarking, not frequency scaling.