The 50% is specifically on percona-tpcc[0]. I got there through a mix of batching (postgres processes a row at a time), prefetching, and several handful of other optimizations.

  [0] https://github.com/Percona-Lab/sysbench-tpcc