It are not only the fastest supercomputers that get more powerful each year. Dongarra's laptop for instance, runs at about 2 Glop/s today. That is roughly the power of a TOP500 supercomputer 15 years ago. And you do not need that if you only do e-mail.

That the performance of supercomputers can be used to solve real problems can be illustrated with the winners of the Gordon Bell Award. This award is given to the real application that runs the highest sustained performance on a supercomputer. The performance of the Gordon Bell price winners follows the same pattern of growth as the high-end TOP500 supercomputers: in 1988 they broke the 1 Gflop/s barrier (yes: today that can run on your laptop, Dongarra said.) Tflop/s in 1998 and 1 Pflop/s in 2008. So we could expect Eflop/s around 2018. By that time a laptop will run at 1 Tflop/s, comparable to the 1998 Gordon Bell supercomputer performance.

But there is still a long way to go before Eflop/s systems can deliver sustained performance to the next Gorden Bell price winner. The main problem is that you cannot scale up machines the way that we are used to, according to Jack Dongarra. Since about 2005, the clock speed of processors has hardly increased. Reason is that if you make computers faster this way, also the power consumption goes up. Processors would heat too much: and it would take too much energy and cooling.

However, the chip companies, like Intel and AMD are still able to squeeze more and more transistors on a chip. But instead of using this to create a faster processor, they put more processors - called cores then - on a chip. This allows performance growth again. But there is one problem, Dongarra said, when more cores are put on a chip, it gets more and more difficult to get the data on and off it fast enough. Also the growth of the on-chip memory does not increase at the same speed as the processor performance. Hence the existing software models to use parallel systems based on message passing techniques do not work anymore. And it still has to be figured out if there is another paradigm that does work. So basically, the hardware vendors that could not solve the problem of getting processors faster, turned their problem into a software problem and handed it over to the software developers.

Jack Dongarra, and many other people, are working on these new software paradigm that includes rethinking and rewriting applications, algorithms and software.