The current fastest is in China, it cost 390 million dollars to build and has 34 petaflops. Obama wants to build an exaflop computer that is thirty times faster.
Economically, with Moores law isn’t depreciation so rapid you can’t recoup your investment? Wouldn’t a computer with speeds comparable to the world’s best only cost 1/8 as much to build within four years? If so wouldn’t you have to recoup the cost to build it within a few years?
Granted you can still use it, but that 390 million dollar computer would have to compete with 100 million dollar models within a few years, then models that cost tens of millions, etc. Within a couple decades home computers could reach those speeds.
Good question! I’m guessing the hope is that your 360 million dollar computer can generate enough uniquely marketable data in the 4 years before it’s outdated that it would recover the costs.
Granted, I have no idea what supercomputers are used for, but if it allows me to invent a product before anyone else can, and that product generates 2 billion dollars in sales for me, then it would have been worth it. Had I waited until the same computer’s price was reduced to 100 million, then someone else would have beaten me to inventing the product.
Being government projects, they don’t have to recoup the whole cost. Simply providing the service is enough of an incentive. The project advances science, which benefits everyone.
Supercomputing is the way to solve certain kinds of problems, usually simulations. They’re really nice in climate sims. They’re also useful in modeling protein behavior at the atomic level.
The stuff I’ve heard faster supercomputers are needed for are better weather predictions and I assume climate predictions. Traditionally there has been a need for them for modeling nuclear explosions, I don’t know if that is still true.
I’m not sure that business and economic information is accurate enough so that doing more processing is going to help in getting better results.
Mostly I think people build bigger and faster supercomputers for bragging rights.
Pay off the national debt via Bitcoin mining.
As I understand it, today’s “supercomputers” are basically thousands of commercially-available microprocessors hooked together with ultra-fast shared memory and running under operating systems that support applications that can take advantage of massively parallel computing.
They’re typically used for applications like weather forecasting (where the finer-grained the simulation you can run, the more accurate the forecast), simulations of nuclear explosions (ditto), and simulations of things like the birth/growth of galaxies or clusters of galaxies (mega ditto).
I don’t believe they’re used much for product development, but I could easily be wrong.
Governments don’t do it for the profit, they do it for the military
There is a difference between high performance supercomputing systems (Mira, K Computer, Tianhe-2, Titan) that are used for climate modeling, protein folding, hydrocode/plasma dynamics simulation, et cetera, and high throughput computing systems used for high energy particle analysis, ‘Big Data’ analytics, et cetera. In the former, performing complex operations and transformations needed to simulate large scale non-linear physical processes requires high performance at levels of problem discretation or abstraction is the focus, and so high performance of individual processes (which themselves might be split between cores or nodes, and have certain portions handled by dedicated math/graphics co-processors) is key, while in the latter case the ability to handle enormous volumes of data (many terabytes) and filter through to identify specific patterns or behaviors is the goal. In this case, throughput speed is more important that processing speed at the individual process level, and the challenge is in developing the analytics that can automate finding patterns and connections in the data.
And yes, the need to have the absolute fastest machine today rather than spending a fraction of the cost to get the same capability in a couple of years is largely bragging rights, and/or being the first to be able to announce a discovery facilitated by access to this capability. Scientists at the top of their fields often have large egos and are driven by a desire to be acknowledged as the first and best…just as in every other field of endeavor. This is, after all, why Goldfinger built that elaborate unfolding diorama of Fort Knox, only to gas the only people who had seen it a few minutes later. We might conclude that many scientists are just a mint julep and Pussy Galore away from being Bond villains. Think about that next time some poindexter starts going on about finding the Higgs boson; in reality, he’s just trying to figure out a way to use these new discoveries to crack into the Hindu Trivandrum and teleport away all of its gold where he can lust over its brilliance and divine heaviness.
Stranger
[nitpick] An exaflops computer. One exaflops = 10^18 floating point operations per second.[/nitpick]
Today’s bragging rights computer is tomorrow’s garden-variety weather simulator. Part of the point of building the computer is to learn about what it takes to build them, and to spur development of the necessary technologies.
My company makes a decent chunk of change building chips for supercomputers. They’re essentially variants on consumer-grade hardware. Even this cost delta wouldn’t be worth it unless someone was footing the bill, but once developed, the technologies can be applied to smaller-scale systems.
There’s also a protectionist argument for them (not that I agree necessarily, but it’s there). All of the top supercomputers make some attempt to use domestic products and employ domestic talent. It’s like agricultural subsidies in that respect.
The incentive to not get pwned at Call of Duty!
This, and science.
For government, cost effectiveness isn’t based on profit margins, it’s based on preservation of the state and mankind - losing either of which would put a pretty big crimp in the budget.