The Need for Speed
When it comes to supercomputers, there is really only one metric: Speed. Computer speeds are the main measuring stick used to compare supercomputers. This is why research and development in the computing field is always geared towards increasing speed, or developing materials that could increase speed.
And when your goal is the fastest supercomputer in the world, you pull out all the stops. A European company, ATOS, has just revealed its new product—the Bull sequana, the world's first exascale-class supercomputer. And it has a tall order: the company hopes the sequana will have exaflop-level performance by launch date in 2020.
The computer is being built for the French Alternative Energies and Atomic Energy Commission (CEA). It has previously built supercomputers for CEA, with the 1.25-Pflop Tera 100 completed in 2010. It was then considered one of the fastest in the world, but is now at 74th.
To the uninitiated, an exaflop is a billion billion floating-point operations per second (flops). Today's fastest supercomputer, China's Tianhe-2, only has a maximum performance of 33.9 petaflops (millions of billions of flops).
Speed and Efficiency
ATOS' goal is not only dramatically increased speed. It also wants to make strides in efficiency, promising to be ten times more efficient than current machines.
A look at the old Tera 100 isn't really that promising. It's a laggard when it comes to energy efficiency, at just 0.23 Gflops/watt. Average power consumption across the top 500 supercomputers was 1.45 Gflops/watt. With that, an exaflop machine would need around 690 MW—that's around a third of the output of the Hoover Dam!
The most efficient machines, most delivered within the last year, perform two or three times better than average, delivering between 3.77 and 4.78 Gflops/watt.
But if ATOS wants to beat the average machine by a factor of ten, it will have to build a supercomputer that's three to four times better than its most efficient predecessors—a performance improvement of 300%.