Slashdot Mirror
US DOE Sets Sights On 300 Petaflop Supercomputer
dcblogs writes U.S. officials Friday announced plans to spend $325 million on two new supercomputers, one of which may eventually be built to support speeds of up to 300 petaflops. The U.S. Department of Energy, the major funder of supercomputers used for scientific research, wants to have the two systems – each with a base speed of 150 petaflops – possibly running by 2017. Going beyond the base speed to reach 300 petaflops will take additional government approvals. If the world stands still, the U.S. may conceivably regain the lead in supercomputing speed from China with these new systems. How adequate this planned investment will look three years from now is a question. Lawmakers weren't reading from the same script as U.S. Energy Secretary Ernest Moniz when it came to assessing the U.S.'s place in the supercomputing world. Moniz said the awards "will ensure the United States retains global leadership in supercomputing." But Rep. Chuck Fleischmann (R-Tenn.) put U.S. leadership in the past tense. "Supercomputing is one of those things that we can step up and lead the world again," he said.
I remember back in the 80's all the excitement about building faster and faster super computers to solve all sorts of grand challenge problems and how a teraflop would just about be nirvana for science. Around 2000 teraflops came and went and then petaflops became the new nirvana for science where we would be able to solve grand challenge problems. Now exaflop is the new nirvana that will solve grand challenge science problems once again. Seems raw computing power hasn't given us the progress in science we predicted. Sure it's been used for stuff, but it hasn't helped us crack nuclear fusion for instance, one of its often hyped goals.
Where's the score card on how much progress has been made because of super computing? I know drug design is one very useful application, but what are other areas that have been transformed?
Letter To Iran
There are plenty of things that can use all the computing power you can throw at it these days. As you mentioned, weather forecasting - though more generally, climate science. Somebody from one of the National Labs mentioned at a college recruiting event that they use their supercomputer for (among other things) making sure that our aging nukes don't explode while just sitting in storage. There are thousands of applications, from particle physics to molecular dynamics to protein folding to drug discovery... Almost any branch of science you can find has some problem that a supercomputer can help solve.
Additionally, it's worth noting that these generally aren't monolithic systems; they can be split into different chunks. One project might need the whole machine to do its computations, but the next job to run after it might only need a quarter - and so four different projects can use the one supercomputer at once. It's not like the smaller computing problems end up wasting the huge size of the supercomputer. After all, many of these installations spend more in electricity bills over the 3- or 5-year lifetime of the computer than they do to install the computer in the first place, so they need to use it efficiently, 24/7.
On the contrary, modern supercomputers are designed for energy and thermal efficiency that rivals and exceeds that of smartphones. Granted, you wouldn't want to put one of these NVidia chipsets in a smart phone, but in terms of compute power per watt, they're far more efficient than general purpose computers.
That said, they do consume a lot of power. But that's precisely why they're engineered for efficiency -- when you're getting the bill for such a monster, that extra 10W/core adds up big time.
I do not fail; I succeed at finding out what does not work.
There are plenty of things that can use all the computing power you can throw at it these days. As you mentioned, weather forecasting - though more generally, climate science. Somebody from one of the National Labs mentioned at a college recruiting event that they use their supercomputer for (among other things) making sure that our aging nukes don't explode while just sitting in storage. There are thousands of applications, from particle physics to molecular dynamics to protein folding to drug discovery... Almost any branch of science you can find has some problem that a supercomputer can help solve.
True enough, the rub is that developing solutions for those problems that effectively use supercomputing resources is as big a problem as the problem. It's more than likely you are reading this on a multiprocessor with a vector acceleration system, that has more potential compute power than any supercomputer from older than 15 years. The question is just what is your utilization and where is the speedup from all the extra compute resources.
You forgot the Tsar Bomba http://en.wikipedia.org/wiki/T...
But then again there were so many. It's kind of mind numbing that we have to borrow stupid from the former soviet union.
P.S. The soviet era is Lenin to breakup, the life of the Soviet Union.
Sorry, the NSA needs all those cycles to process everyone's phone calls. Remember, it's only illegal for a person to listen in on your calls.
Sleep your way to a whiter smile...date a dentist!
The number of floating point operations (FLOPS) performed by a next-generation game console outranks early days supercomputers like the Cray.
Sure, but do they have the system capability / bandwidth to actually do anything with those numbers and is their raw speed offset by not being vector processors like the Cray 2 (process an entire array of data in 1 instruction)? I'm not a hardware geek, but was an administrator for the Cray 2 at the NASA Langley Research Center back in the mid 1980s and, among other things, wrote a proof-of-concept program in C to perform Fast Fourier transforms on wind tunnel data in near real time - probably would have been faster had I been a FORTRAN geek - and the system could pump through quite a bit of data - at least for the 80s.
And the Cray 2 was way prettier than a PS3/4 or Xbox, though the Fluorinert immersion used for cooling is a bit cumbersome and expensive :-)
It must have been something you assimilated. . . .
...also there is the inertia of so many scientists and engineers...
Sounds like words of a youngster who doesn't know that newer isn't always better.
It must have been something you assimilated. . . .
Well, you could just actually test old and unrefurbished nukes to see just what all those decay products accumulating beneath their shells do, or you could just simulate it. No wait, the politicians have sworn off all actual testing, you can only simulate. Back in the 2000's Supercomputers were all we had to tell us what was in the decomissioned former Soviet nukes they were asking us to open up and get the Plutonium out of - some were seven to ten years behind scheduled maintenance and nobody was sure just what had built up in it, but the Russians still had Chernobyl in their minds and would love to comply with the treaty by destroying it, it was just their technicians were getting readings as soon as they opened up the outer casings that convinced them they would have died if they had gone any further.
It's no accident that most of the US title holders for fastest supercomputer have been built at the Oak Ridge National Laboratory. The whole US supercomuting program to date has cost much less than one decay induced explosion releasing the sort of stew of Polonium, Americium, and other incredibly virulently radioactive glop that builds up in old nukes, simply because all the possible scenarios are so ultimately nasty, as in covering the area of 100 Chernobyl's nasty.
Who is John Cabal?
Not sure what point you're trying to make here, but newer supercomputers are very different from those early supercomputers, in far more ways than one. The parallelism is much higher (supercomputers now have millions of nodes, with exascale computers expected to have tens of millions or more), for instance. It's extremely hard to program for them. Interconnects have not been improving very much and so data flow between cores has to be managed carefully.
A fool and his hard drive are soon parted.
Sure, but do they have the system capability / bandwidth to actually do anything with those numbers and is their raw speed offset by not being vector processors like the Cray 2 (process an entire array of data in 1 instruction)?
Nope. The vetor unit with its crazy chaining and entire array computations initiated by a single instruction were the tricks required to get the CRAY to be as fast as it was. With all those tricks, the CRAY-2 peaked at about 2GFlops or so. Bear in mind the relative of Vector processing (SIMD) is now present on all high performance CPUs.
The other problem was indeed memory bandwidth. Cray solved this with dedicated processors for aranging memory transfers, multiple memory channels and other tricks. These tricks are now present in modern high performance processors, though the memory co-processors are now built in and not separate or even turing complete processors in their own right.
The clock speed was actually quite low, because the machine was physically large and the speed of light limited what could be done.
There's not much contest now because the hardware has advanced fast. Even early gen Atom CPUs could reach multiple GFlops on benchmarks (as opposed to the 2GFlops theoretical peak of the Cray 2).
But yes, the Cray computers basically looked cooler than any other computer before or since.
SJW n. One who posts facts.
These systems will use IBM Power CPUs and Nvidia's Volta GPU, the name of a chip still in development.
The discussion turns to Kim Kardashian? On Slashdot?!
When you're dead, you don't know you're dead. It only affects the people around you. Same thing when you're stupid.
With tens of millions of nodes data logistics pretty much always is a problem, even for supposedly embarrassingly parallel problems. Either the nodes communicate with only a few neighbours, in which case you have to carefully design the layout of the computations to make sure every node can communicate efficiently with its neighbours, and there probably is also some kind of global clock that has to be maintained. Alternatively you have some kind of farmer-worker setup where each worker node is happily chomping on an problem on its own. Even then you have to have farmer nodes that keep all those millions of little chompers busy. That is usually a headache on its own, because they will need some data to get started, they'll report back some data, and that's a lot of data if you deal with so many nodes.
If all those millions of nodes need to consult some kind of global data, even if it is rarely, that's another data logistics headache. And those are the best-case scenarios, and that's ignoring any fault-tolerance issues, which with tens of millions of nodes is already far into the `happy fool' area.
So yes, it is extremely hard to program for such an architecture. The only alternative is to use a middleware such as Hadoop where you try to fit your problem into a certain computation pattern (`skeleton' was a popular term for this for a while), and let the authors of the middleware worry about all the headaches I mention above. That doesn't mean the problems aren't there any more, it is just that the middleware authors are trying to hide the issues from you as well as they can.