This recent article from Big Data Republic describes a hands-on performance comparison between GraphChi - a graph computation tool from GraphLab - and a 1,636-node Hadoop cluster. The task set for both was to process a Twitter graph with 1.5 billion edges, and the result, surprisingly enough, was a significantly quicker processing time for GraphChi.
The article goes into detail on the techniques behind the optimization of GraphChi, and seems not so much to point to the superiority of GraphChi (though it does lean that way), but to highlight the importance of planning and optimization when it comes to massive datasets. It's an interesting look at the work that still needs to go into big data tools to make them as efficient as they can be.
Take a look at the full article for a look at how much difference optimization can make. Also, to be fair, this comparison originally took place a year ago. Hadoop 2.0 rematch, anyone?