2013: it's time to stop chasing flops

From the translator: Exascale computing is such an ambitious project to achieve ExaFLOPS performance by 2018. It is believed that high-tech computing is already closely in petaflops. Is it really? Reflections on this topic by William Gropp, director of the Parallel Computing Institute, were published in The Exascale Report .

2013: TIME TO STOP TALKING ABOUT EXASCALE
William D. Gropp

Everyone who reads this believes in the power of computing technology. It seems to us self-evident that the performance of the most powerful computing systems must continue to grow at the same speed in order to meet the needs of society. However, it is not so indisputable.

Worse, the emphasis on any measure of performance (not necessarily ExaFLOPS) instead of focusing on the ability to solve immediate problems can cause (and forces!) Us to focus on technology, not what can be achieved with this technology. In turn, such projects as Exascale draw funding from, say, projects in Big Data .
')
So, if we stop talking about performance, then what is worth talking about? For me, this is obvious: we must discuss and describe the challenges and opportunities that we face, from basic science to commercial solutions, and only then - the important role that high-performance systems play in solving these problems. The problem itself must be clearly defined before its computational complexity, and the only way to show the need for faster computers.

Too much effort today is attached to the pursuit of flops, while you first need to ask: what can we do with them? The danger is that it suggests approaches that really don't need exaflops - simplified, non-optimal algorithms or distributed research. It will do more harm than good, and even a layman can easily debunk the claim that such computing power is needed.

Don't misunderstand me, I believe that we really need much more powerful systems to solve the problems we face, whether it is understanding the functioning of life and the universe or developing sustainable infrastructures. But the need for them must clearly come from the problems.

We can start to move away from the emphasis on FLOPS (especially if this is the result of benchmarks that can be misleading, although the largest CS communities take them seriously) and focus on solving the most difficult computational problems. Among other things, it provides the best rationale for developing new technologies that are needed to create ever faster machines, which we all believe are so necessary. Yes, without such a simple metric as ExaFLOPS, it will be more difficult to quantify the result, but no one will argue that a high-performance system cannot be described by a single number.

In order for new record performance to become a reality, we must stop chasing performance.