In this new whitepaper, authors Marc Snir, William Gropp, and Peter Kogge describe the showstoppers faced by fellow scientists seeking to reach sustained Exascale performance in the coming decade:
We could, conceivably, push supercomputing performance by brute force: Build systems that consume 500MW or 1 GW and require machine rooms (or machine fields) the size of tens of football fields. Even if the cost of such an approach is acceptable, it is not clear that brute force can lead us to exascale: large physical size is an impediment to scalability and a source of frequent failures (failure rates today for Blue Gene systems have been reported on the order of 0.001 failures per year per socket, and projecting forward to a million sockets yields socket failure rates of minutes). Furthermore, with such an approach, the US will forfeit any advantage over competitors and possible enemies that are capable and willing to outspend us.
Download the whitepaper (PDF).