As every year at this time we get the release of the Top 500 world's fastest supercomputer. Since April 2014 the fastest supercomputer was Thiane-2, a Chinese supercomputer whose engine (microprocessors) were Intel chips (plus a Chinese chip used for interconnection). It crunched at a rate of 33 PFLOPS (33 million billion floating instructions per second) and remained for these two years at the top of the ranking.
Now, in the just published Top 500 ranking, June 20, 2016, we have a new number one: Sunway TaihuLight crunching three times as much, reaching 93 PFLOPS. And this one is 100% Chinese. All its chips are designed and made in China. The Sunway is a cluster of 10,649,600 processors cores organised in 40,960 nodes with a 1,31 PB of memory. These nodes are interconnected one another with a 96Gbit per second lines creating a broadband highway at 70TBytes per second (now, this is broadband!). The total power consumption is in the order of 15MW (for comparison the FrecciaRossa 1000, the fasted train in Italy has a power consumption of 9.9MW).
Of course, the big question mark with supercomputers today is the huge parallel computational architecture. Unless you have an application that can be run in massive parallelisms you cannot leverage on the supercomputer massive capacity. IN the case of TaihuLight the researchers in Wuxi have announced that three of the applications they are running on it make use of 8 out of the 10 millions processors, pretty good indeed!
What does this processing power means? To put this into perspective, this processing capacity should be similar to the ones of Google global capacity. This was estimated to be in the order of 40PFLOPS in 2012, hence assuming a doubling of that capacity in 4 years makes sense. I was not able to find more recent estimate, if you have one please let me know. Notice, however, that measuring processing capacity is a bit elusive, since it really depends what you are measuring. As an example, one can measure the specific capacity to manage transactions rather than processing instructions. In this area Bitcoin is probably ranking number 1, with a processing power in terms of transaction management (hashing) that is 100 times higher than the one of Google.
TaihuLight is a big blow to the US supremacy, a first since the creation of the Top 500 list back in 1993.
An even bigger blow if we consider that till 2001 there was no Chinese supercomputer that made the list of the 500 most performing supercomputers. In the June 2016 ranking out of 500 best supercomputers 167 are in China, 165 in the US. Just 4 years of declining investment in supercomputers on the US side have let China take the lead.
US has launched the National Strategic Supercomputer Initiative that aims at the Exaflop computer and will deliver in 2017/2018 a 200 PFLOPS computer (twice the performance of Sunway TaihuLight), but of course one has to see what China will come up with in the coming years. The NSSI aims at delivering an Exaflop computer by 2023, whilst several observers believe that Cina might reach that mark in 2020.
IEEE - FDC has launched three years ago the Rebooting Computing initiative to tackle the increase of performance starting from scratch, through a complete redesign of computer architectures and new OS paradigms. It won't deliver Exascale computation and time soon, but is something that might result within the next decade in a fundamental change in computation, bringing supercomputing power in everybody's hands in the next decade closing the time gap between supercomputer performance and mass market processing performance (that is now around 15/20 years: it takes 15 years of evolution to bring the processing power of a supercomputer to the mass market. Our cell phones pack the 1990ies supercomputer processing power).