There are certain substantial steps you have to make to reach exascale because if you were to just extrapolate today's technology, the power consumption of an exascale machine would be huge, too large to make it practical.
You could put an exascale system together today if you wanted to, you'd just take thousands or 500 of the TOP500 list. If you have a large enough space and enough electrical power, it will probably run.
But you have to fight with resiliency While you already have certain levels of failure rates in petaflop systems today, in exascale systems they will be even higher.
The very underlying thing is energy-efficient transistors, because the foundation element of any compute structure is a transistor.
Also memory technology needs to change. How do you interface from memory into the processing element, because you have to shuffle in and shuffle out data very fast if you want to be practical in an exascale application.
So those metrics are not unknown to us because to reach a petaflop you already had to something there. If you look back in time with our CPUs, we have kept the power envelope roughly the same over a very long period, but the capabilities of those CPUs with a higher number of cores and better cores and better instruction sets... have dramatically improved.
So it's not an unknown that you can improve performance whilst staying the same.
Have you any predictions on timeframe for when exascale computing will become a real-world possibility?
Well I wouldn't dare to say there are any realistic ones. There are experts who claim to know when it will happen, but I wouldn't claim any year.
It's not ages away but it's not absolutely near term.
