Slashdot Mirror
Sandia Wants To Build Exaflop Computer
Dan100 brings us an announcement that Sandia and Oak Ridge National Laboratories are setting their sights on an exaflop supercomputer. Researchers from the two laboratories jointly launched the Institute for Advanced Architectures to facilitate development. One of the problems they hope to solve is how to provide each core of each processor with enough data so that cycles aren't going to waste.
"The idea behind the institute — under consideration for a year and a half prior to its opening — is 'to close critical gaps between theoretical peak performance and actual performance on current supercomputers,' says Sandia project lead Sudip Dosanjh. 'We believe this can be done by developing novel and innovative computer architectures.' The institute is funded in FY08 by congressional mandate at $7.4 million."
Aren't we getting a little bit ahead of ourselves, Sandia? What program would you run on this? This brings up the essential issue: what kind of program would YOU write to take advantage of this? I can only think of one: AI.
Seems to me that an SSD 4GB Fibre SAN might be such a novel way to go to ensuring that each node gets pretty much continuous access to large data. Seem to me the throughput would run closer to peak all the time as opposed to using a traditional HDD-based SAN. Combine that with some sort of clustering technique and I think you could achieve really good performance.
My blog
As long as they keep making new peripheral buses and networks that are the fastest in the world to keep up with the supercomputing speeds, that we can then buy to use with our PCs, then it's a great investment.
But since the American people will have bought the new tech for the world, it should be released into the public domain, after maybe 5 years patent licensed to only American corporations (who cannot sublicense it abroad). That's what investments in American tech should be like. Not just subsidies to private corporations, however Chinese they might be.
--
make install -not war
However, at the moment there are no serious applications that will only become feasible by having more computer power.
More speed in calculation has plenty of benefits, but AI as a research field will not be making major announcements soon because of this new machine.
How long does it typically take for memory and sotrage advances to make to end consumers? For example, when we first heard about "gigabytes" back in the day, how long did it take to get there once it was being done in the laboratory?
OK, here's the truth. I'm just wondering since I need more memory to carrying around the entire internet in my pocket. Right now, I can only fit Ron Paul fanatic postings on my USB stick. They are taking up a lot of room. (Nothing against Ron Paul, mind you.)
How to Download YouTube Videos
it said Santa was building a super computer :)
This sounds a very interesting OR optimization problem, but I am not sure what are the variables...If a processor is working on a particular piece of a problem and the data required to solve this will be made available by some other processor located away from the processor, I guess we are really talking wait times and not distance, am I right?
No cycles wasted there!
I don't know why, but I want one.
Twenty years ago we had a Compaq portable that ran on a 16 mhz 286 at work, and it was HOT. Blazingly fast, could do anything. That is, for its time. The supercomputers then weren't as powerful as your laptop today.
So if I can manage to stay alive for another 20 years, I'll probably have a laptop more powerful than the supercomputer in TFA. I guess I'll just have to wait a while.
-mcgrew (link is to "Growing Up With Computers", a 2 year old K5 article)
mcgrew's razor: Never attribute to stupidity that which can be explained by greedy self-interest
7 million isn't alot for a datacenter with a supercomputer housing a novel architecture. They'll need infiniband or gig fiber, and other high end equipment. That in itself will take 2 million at least. I dunno. Maybe they'll do a low-rent google and call it unique.
It'd be cool if TFA's headline was actually correct: Then we'd have a machine whose performance actually accelerated by a 10^15 floating point operations per second *per second*. That gets to be a lot of FLOPS real fast.
OTOH, it might just be the singularity happening. We wouldn't notice until it was too late.
...Skynet?
Hope its not a big "flop"...
... the first computer with enough specs to able to run Vista! (just)
Reading the article, the goal is nowhere near building a real exaflop computer, but more about thinking about issues (like processor data feeding).
In a year and a half, we shouln't have more than 100 GFlops per socket, which means that you will still need 10 millions of processors (not cores!) to achieve the exaflop computer. No chance to build a cluster that big (at least these years).
The all-times progression of the top500 shows that exaflop computers should arrive around year 2020, definetly not tomorrow. (x10 every ~4 years, 2008:1 PF, 2012:10 PF, 2016:100 PF, 2020:1 EF)
Firstly there are a whole lot of applications that can be thought of which would be able to use this ... for example
dynamic data mining like image search with face recognition or virtualisation like virtaulisation of the entire human body to perform heart surgery in virtual environment as an learning experience for budding surgeons, also run time perspective rendering of videos, etc. such applications can be written with current programming capabilities. I don't doubt that they will be very crude and also inconsidered of time and space complexities optimisation.
the optimisation cud be achieved later when paralleling programming evolves, and global standardisations are made and conventions set for programming as well as architecture design n development.
I'd run me. Kurzweil suggests we run around 12 petaflops or so. The problem is getting the data out of my head. I'm not letting them slice my brain up until they figure out how to scan the synaptic weightings in addition to the connections. Histography, confocal laser scanning, electron microscopy, those seem to be the holdups to uploading now, not processing time.
refactor the law, its bloated, confusing and unmaintainable.
obSlashdot: but will it run linux?
RS
Shoes for Industry. Shoes for the Dead.
For instance the Cell/BE processor allows C/C++ programmers to manage memory directly with the Memory Flow Controller to perform double-buffered asynchronous transfer of data between the main memory and the processor memory. Using Direct Memory Access, Cell users can achieve 98% of the peak performance on some applications: http://www-128.ibm.com/developerworks/power/library/pa-cellperf/ .
But can it do realtime unbiased raytracing? As a CG artist, I would love to sit down at one of these things and let it hammer away at my scenes with photorealistic quality at 100fps
Duke Nukem Forever
It's not hard to come up with programs that need a lot of processing power to run. Most of the stuff currently being run on supercomputers are (relatively) small programs, but with huge data sets that are easily parallelized. Pretty much any kind of simulation falls into this category: climate, genetics, biology/pharmaceutical, plasma physics, particle physics, nuclear physics.
You don't need a "smart" program to utilize a fast computer - in fact, they are most useful in situations where the smartest people in the world can't figure out algorithms to do things any faster, and there is no alternative but massive grunt-force calculation.
will it play doom?
A lot of so-called supercomputers only have some parts that may run at the theoretical speed, because they stint in other parts such as memory, or bus speed etc. A viable general-purpose computer usually has one flop = one byte of core = one second to completely write core. Thats pretty much the case with desktops in the single-gigaflop range.
True. Sandia does a lot of nuclear systems simulation.
... the list goes on, but almost any US researcher can get an account on their systems, and purchase cpu time. The trouble is that neither of these labs has quite enough resources to dramatically change computer architecture directions. They can both afford to have 1 or 2 very high end machines at a time, which is enough for companies like IBM or Cray to make subtle changes in their product lineup, but not enough to completely redraw the map of the HPTC marketplace.
However, Oak Ridge is an unclassified facility doing mostly academic research on climate change, fusion energy, biological systems modeling, geological systems,
Imagine a Beowulf cluster of these .... oh nevermind
Sandia and Oak Ridge are not coming up with an exaflops computer. They are contemplating how to write software, in a way that will effectively use exaflop computers when they become available. This little group has a budget of $4.7million, which is enough to pay a dozen high-level research scientists, and a half dozen software developers for a year. They're not going to reinvent high performance technical computing.
They're going to rework some fundamental math libraries to deal with the obvious trend in HPTC. They know that they are going to have to learn to deal with a million nodes, each with dozens or hundreds of CPU cores. The ratio of memory latency to processor speed is getting worse, not better. The ratio of interconnect bandwidth to node performance is getting worse, not better. Memory capacity per node is going up, but capacity per processor core is flat or downward. Disk performance per flop is flat or down. Checkpoint/restart of a million nodes in a non-starter. Flops are more plentiful, and massively cheaper, with each generation, but there are costs associated with that. There are a lot of paradigm changes in the way that these systems are managed, and the way that these codes are written, to deal with the limits imposed by these machines.
Vista of course! It might actually run okay on that hardware.
Your thin skin doesn't make me a troll
Well I'm all for it, as long as they store it on hard ware that I can take home with me.
Has the "Just imagine a Beowulf cluster of these" joke gone out of fashion? In the old days, someone would have posted it before the ink was dry on the story (and virtual ink dries really fast).
1. Build exaflop computer.
2. ???
3. Profit!
...laura
Chapter 4 of Getting Up to Speed: The Future of Supercomputing (http://www.nap.edu/catalog.php?record_id=11148#toc) outlines plenty of potential applications.
I've recently seen a Presentation on the issues around exascale. The end result ... EVERYTHING must fundamentally change. The power requirements would alone swallow the needs of a small town. Even DDR4's bandwidth would be a second rate solution. The scheduling requirement across the hundreds of thousands of processors could consume the system alone, much less allow it to work. Only one of the supercomputing giants is actually building systems (for five years out from now) that will scale out to past 5 petaflops and Sandia is thinking about 1,000 petaflops.
The linux weenies might as well forget it because there's no way in hell that linux could ever run a box at Exascale. The hardware alone will be more than 12 years out.
Given that the war in Iraq is costing US taxpayers about $2 Billion per week (pre-interest), Sandias program is budgeted at roughly 40 minutes worth of the direct war effort, and given that new approaches to distributed and other forms of HPC will come out of this, with some real-world application (and ROI), this is relatively high value. Hopefully, Slashdot readers will also read some of the technical papers published as a result of this work, and bring some real value out of it.
One Supercomputer ring to rule them all, One Token-Ring to find them, One Ring to bring them all, and in the LAN bind them. In the Land of Ethernet where the switches lie.
GCS/MU/P d- s:- a-- C++++$ UL++ P+ L++ E+ W++ N o K- w--- O M+ V- PS+++ PE Y+ PGP t+ 5- X R++ tv+ b++ DI++ D++ G+ e++ h-
I had a chance to listen to talk by a guy involved in this just the other day. You're right about the first part; it is more about thinking. He said a big problem now is that, while they greatly appreciate being funded to build faster computers, there isn't enough storage space and especially *bandwidth* to deal with it!!! In other words, our networking infrastructure in the United States needs a massive upgrade, and so far no one seems to care. Congress will fund a new supercomputer every few years, but for some reason they don't see the need to invest in converting to fiber optics, for example. Part of the point of this project is to make people aware that we need to create infrastructure to handle all of this data flow if we want to progress.