My Supercomputer is Bigger Than Yours!
Once again, China officially has the fastest supercomputer in the world. All the careful preparations of marketing departments throughout the HPC community leading up to ISC were rendered mute, as the usual slurry of ISC’13 “me too” press releases were blown aside by the revelation of Tianhe-2’s 50+ petaFLOPS.
Chinese supercomputing had again sprung from rumors to deliver the harsh reality that the USA was no longer home to the most powerful known supercomputer in the world. China’s new supercomputer is not only faster than the USA’s leading contenders – it is twice as fast. And, to pour salt into the American wounds, Tianhe-2 is not a stunt machine with buckets of cheap FLOPS lashed together with just enough wet string to run Linpack. It is a custom designed supercomputer combining next generation Chinese interconnect technology with American CPUs and HPC coprocessors.
The USA’s HPC community seems unsure whether to hide under the duvet and try to reassure themselves of American HPC leadership by quoting other metrics, or to seize upon this opportunity to demonstrate to their government masters how other nations are aggressively pursuing supercomputing and thus focus their efforts on securing funding for exascale and other future HPC needs. Meanwhile, the European HPC community enjoys a hint of smugness that the USA’s leadership has been taken away, smothered by an uncomfortable knowledge that such a feat is unlikely to ever be achieved by Europe.
Having a more powerful supercomputer is not merely useful for “mine is bigger than yours” contests – a more powerful supercomputer can deliver more science and engineering in a given time than a smaller system through sheer capacity. It can enable major advances in science and engineering through capability – exploring the leading edge of what is possible with modeling and simulation at scale. It can inspire a generation of users to pursue computational science and engineering. It can inspire a computing technology industry and wider commercial applications of HPC.
Indeed, a more powerful supercomputer is so important that nearly everyone who has a supercomputer tries to find criteria such that theirs is the leading system in a given category, whether “fastest commercial system”, “biggest academically owned system”, or whatever.
Yes, size matters.
But – what if size did not matter? Pretend that all supercomputers were the same size and couldn’t be made bigger. Or perhaps they were all so big and cheap that any user could get as much resource as they needed with zero wait.
In this obtuse reality, the size of the supercomputer no longer correlates to the capacity or capability of science that can be achieved.
What would matter? Other parts of the ecosystem would become the enablers of computational leadership, to produce the leading edge science and engineering, and the resulting economic benefits. Software, people, applications, etc. would become the differentiators.
The researchers who could lead the way in a given computational field would no longer be the ones who had access to the biggest machine, but the ones who could make best use of the same machine size as everyone else. That might mean the most scalable code, or the fastest code for a given problem size, or the most robust and accurate code. It might mean the group who had the best skills strategy to ensure continued development of the computational skills within their group.
Companies could not secure a competitive advantage through computing capacity – they would have to seek better algorithms (scalable, accurate, validated, …) and better investment in the people (developers and users) who could turn computational applications into business results.
How might today’s familiar international competitiveness arguments change in this weird world? There would be no point urging governments to fund development of technology (our pretend world assumes hardware can never be the differentiator). Evoking national pride by deploying bigger systems than rival countries would be impossible. The national need to pursue competitiveness could only be serviced by supporting the development of algorithms, computational methods, scalable software engineering, scientific applications, etc. – and above all a pipeline of computationally aware people/skills.
How would a Top500 equivalent work in this world? (Because there would still be a natural human need to measure progress and compare with other computational users.) I have no answer to this – but it is probably a critical question (even in the real world, not just my pretend world) – how to measure supercomputing capability if not by anything directly correlated to size of the machine?
Instead of tracking roadmaps from hardware vendors, technology planning might consist more of algorithm roadmaps, software implementation roadmaps, recruitment & mentoring proposals, etc.
Ultimately, nothing in the ranking of supercomputing players would change – the richest countries and companies would still be the winners as they could invest more strongly in people (basic methods research, software engineering, science applications, etc.). Some countries/companies would “punch above their weight” – those who understood the need to invest in the right things and did so with more commitment than their rivals. (How is that different to the real world?)
Indeed, perhaps that is where my little “size doesn’t matter” experiment leads me – to conclude that the leadership in supercomputing (and thus the benefits to research, innovation, economic impact, etc.) will always belong to those who understand what supercomputing can do, along with how to do it better – and then act on that understanding.