I suppose he's going for the 16 cores instead of 8 given the interconnect isn't used. It would be interesting to see a performance comparison between the xlarges and the cluster instances on simple speedups like this.