Those Bitcoin Unlimited guys measured the mempool performance and found out it was not the protocol issue, it is protocol implementation issue. Currently transactions commit to the mempool in the single thread exclusively, so it doesn't matter if you have 4 core, 8 core or 32 core processor once the mempool gets saturated it falls apart completely, not being able to process new transactions. They found about this empirically, through watching measurement charts. There is a video from Stanford conference held a week ago where they explained the details, link to Youtube is posted in this thread few pages ago. They wrote a multi-threaded implementation and then (surprise, surprise) made some additional "improvements" on the process, so even if their code will be opensourced it can not be patched directly to Bitcoin Core codebase, developers will have to do some of they own work on it. I'm surprised no one cached this long time ago.