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.
Reminds me a little of Soviet era build the biggest thing you can projects. I could see it if they have a particular problem that either needs faster updates or higher resolution updates than the performance currently available provides (weather forecasting comes to mind). But building big to build big ? The interesting part of high performance computing is all in architecture and software to make use of it. This strikes as a little wasteful
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
Let us see what else are in the past tense ...
How many of the microwave ovens / teevee sets are being made inside the United States of America?
How many of the American jobs have been "outsourced" to places like India or the Philippines?
How many top American companies are using foreigners as their CEOs? ... and the list continues ...
Muchas Gracias, Señor Edward Snowden !
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
had to adjust mine.
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.
unless your the Department of Defense, you still need to worry about your budget. Especially in times like this.
Nowadays you inject silicone into the pig's hips and ass.
(-1: Post disagrees with my already-settled worldview) is not a valid mod option.
lol.. I love that series.
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 subject implies that the NSA publicizes the capabilities of their rigs. I would be willing to bet they have near the computing power of China all by themselves.
For 20+ years, HPC systems have relied on the same conservative design of compute separated from storage, connected by Infiniband. Hadoop kind of shook up the HPC world with its introduction of data locality, especially as scientific use cases have involved larger data sets that distributed data storage is well-suited for. The HPC world has been wondering aloud how best and when to start incorporating local data storage for each node. Summit introduces some modest 800GB non-volatile storage per node for caching (which they call a "Burst Buffer"), but no bulk data storage.
I blogged about how the Summit design seems very conservative, especially for a system to be delivered in 2018, and especially for a supercomputer that is billed to be the most powerful in the U.S. if not the world.
I hate this attitude that if you don't have the top spot, you are crap. It is so silly the attitude that the US somehow lost something by not having the first spot on the top 500 list.
I mean for one thing, the Chinese computer is more specialized than the big US supercomputers. It gets its performance using Intel Xeon Phi GPGPU type processors. Nothing wrong with hat but they are vector processors hanging off the PCIe bus. They work a lot like graphics cards. There are problems that they are very fast at, linpack (which is what's used to test) being one, but others they are not as fast at. Many of the US supercomputers (like BlueGene/Q) use just standard CPUs, meaning their performance holds steady over more kinds of tasks.
Then there's the fact that while the US might not have the #1 spot they have the #2, 3, 5, 7, 9, and 10 spots. In other words, half of the top 10 computers. That is more impressive than having one really big system. Ya it's nice to have a huge system and some simulations need really big systems to do, but there's something to be said for lots of different research groups having access to high power computers.
Also there's the fact that linkpack isn't necessarily the best benchmark.
I'm happy that the US is looking to invest more in HPC because money spent on research is always well spent in my opinion. However let's stop pretending like it is some major failure that the US doesn't have the #1 computer. Big deal.
If the world stands still, the U.S. may conceivably regain the lead in supercomputing speed from China with these new systems
It's kind of hard to regain something you didn't truly lose to China.
Twitter supports and protects racists - by smearing their critics with the "Hate Speech" label.
"Supercomputing is one of those things that we can step up and lead the world again,"
Here's a thought. 100% of chip making companies that make chips that are actually good/fast are American companies. So just don't sell to any other countries in bulk for supercomputer use. Win by cutting off the supply.
A factor of 10 is pretty meaningless in supercomputing. Software quality makes much more of a difference. Of course, politicians are not mentally equipped to understand that and instead want "the larger number" like the most stupid noob PC buyer.
Most ACs are not even worth the keystrokes to insult them. Be generically insulted by this and ignored otherwise.
So, they make this announcement right before the new Top500 list is unveiled in the SuperComputing conference... What clearly means that once again there will be no US system in the Top1 position, right?
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.
Chess is "easy" compared to Go. While Chess requires more bits to store the board the search space for Go is **exponentially* larger. i.e. A a single state of the go board is 2^(19*19) = 2^361 positions = 46 bytes.
Links of interest:
* http://codegolf.stackexchange....
* http://en.wikipedia.org/wiki/B...
The comment above was mine.
The example above is applying the same transformation to a very large number of datasets and then after some hours or days each node writes out what it has done to some shared storage. In that case the "extremely hard to program" thing does not exist since a shell script or queueing system does the job - which is why there are a class of problems known as "embarrassingly parallel". It's not "millions of nodes" but it could be since the problem can be neatly divided into millions of independant parts that are taken on by whatever nodes are available over time. It's storage bottlenecks instead of scheduling that would be the problem with increasing scale.
Some other stuff that is more dependant on what other nodes have done, such as FEA (finite element analysis), can use an interative process where the nodes are fully independant at step1, then the results are taken into account and the job reassigned with step2 with the altered data and so on until there is a good enough solution - the sort of stuff that's been done with computers since before the 1970s where a problem is divided into chunks with inputs and outputs in each element to cut down on complexity. You don't have to immediately know what the nodes working on adjacent parts have worked out until the next interation, which could be hours.
So in each stage in such cases you really only need to know if the job finished correctly. Having storage available to a lot of nodes at once is going to get harder with scale, but actual job management with "embarrassingly parallel" tasks is not a big deal with MPI, PBS/torque or even plain old ssh.
There are of course different problems where nodes do have to communicate with each other a lot but they are not "embarrassingly parallel". As the name suggests it's actually easy to deal with "embarrassingly parallel" problems since the answer looked for is really just a collection of results from a lot of small problems concatenated together in a useful order. A gap in the data at point C doesn't matter if you are currently interested in point D. It's the low hanging fruit of high performance computing but there are a lot of problems that it can solve.