The next big goal for supercomputer designers and builders is to cross the exaFLOP barrier. Original projections pointed to 2018 as the year we would see the first systems, but China decided to step on the accelerator a bit, announcing that it will finish building its new exascale platform before the end of the year. Technically, the supercomputer won't go into action until 2020, but if it keeps this pace, the Red Giant of the East will be the first country to pierce a historic ceiling for the world of computing.
With the Top500 list completely stagnant and the number one supercomputer deploying three times the average processing power of its immediate follower, not a few people anticipate major delays for the creation of the first exascale system. Under extreme conditions, the Chinese TaihuLight can deliver 125.4 petaFLOPS. If my memory and math don't fail me, an exaFLOP is a thousand teraFLOPS, so it would be necessary to multiply the TaihuLight's performance by eight. Difficult? Very. Impossible? Of course not. In fact, it is China itself that is working at full speed to finish its next flagship supercomputer, and if we go by the latest official announcement, the prototype will be ready before the end of the year.
The Xinhua agency publication indicates that according to Zhang Ting from the National Supercomputer Center in Tianjin, this platform will be 200 times more powerful than the Tianhe-1 (the Tianhe-1A revision reached a peak of 4.7 petaFLOPS in 2010), although there are no additional technical details. On one hand, it is speculated that the new supercomputer will be named Tianhe-2A, but any idea of expanding or updating the old system is ruled out due to the U.S. embargo. On the other hand, this exascale supercomputer would not adopt the SW26010 Sunway architecture processors used in the TaihuLight. The name most frequently repeated is the Matrix2000 GPDSP accelerator.
Now, one thing is to finish the prototype, and quite another for it to go into operation. For that, it will need three additional years, placing its formal debut sometime in 2020. This detail gives a little margin to U.S. initiatives, however, their own exascale supercomputer projections go even further, 2023 at best.
