Slashdot Mirror
China Building a 100-petaflop Supercomputer Using Domestic Processors
concealment writes "As the U.S. launched what's expected to be the world's fastest supercomputer at 20 petaflops, China is building a machine that is intended to be five times faster when it is deployed in 2015. China's Tianhe-2 supercomputer will run at 100 petaflops (quadrillion floating-point calculations per second), according to the Guangzhou Supercomputing Center, where the machine will be housed. Tianhe-2 could help keep China competitive with the future supercomputers of other countries, as industry experts estimate machines will start reaching 1,000-petaflop performance by 2018."
And, naturally, it's planned to use a domestically developed MIPS processor
They're using domestically built copies of MIPS processors they copied from someone else (usually wrongly), stringing them together and proving that 2+2=5.
Because that's all America does anymore.
But only Red Flag Linux
A domestically developed MIPS processor?
I wonder how much of that "domestic development" was looking at what foreign companies were making in their country and surrounding areas?
Has anybody even been working on a MIPS processor that they could outsource, for anything larger than little router boxes and things, since DEC died a horrible death ages ago?
since DEC died a horrible death ages ago?
What has DEC got to do with MIPS except that they used them in pre historic, or possibly paleological times. SGI worked on MIPS for years after DEC abandoned them. The lates generation R??000k were pretty snappy and scaled up very well.
SJW n. One who posts facts.
You're right, but it looks like they've done the latter. http://laotsao.wordpress.com/2011/10/29/sw1600-and-alpha-21164/
It says right in the ShenWei Wikipedia article that it's based on the DEC Alpha but something that strikes me as curious is that your article refers to a chip that lasted from 1995 to 1998. So I am to believe that by outright copying a fifteen year old chip from a processor line that has been extinct for a decade or more has yielded a modern day competitive multiprocessing chip?
You can convince me they copied DEC's work. You can convince me they violated IP laws. You can convince me that it is their societal norm to ignore restrictive IP laws. Hell, I'll tell you that right now. But to say that they are doing no work to build on top of these chips feels like it must be erroneous unless what we see is 1990s technology in the ShenWei processors.
This isn't a black and white scenario here. Yes, it's bad that IP laws have been violated. Yes, it's bad that DEC won't see a dime from any of their work being used. But it is also a good thing to have a competitive architecture arise in the world of computing and also it feels good to have a race with other countries for computing power. I can only hope our super computing budget is considered part of the onerous "defense budget" and our leaders who are concerned with a dick measuring contest can dump tons of money into supercomputers for modeling and simulation to scientists while at the same time being able to give the hallowed talking point of "I increased defense spending."
You can start with someone else's good idea, turn it into a great idea and share some credit, right?
My work here is dung.
Not even close.
You'll be lucky to simulate a gram of matter to any significant accuracy, and the higher-level simulations (neural networks, etc.) are seriously lacking in their ability, and to model how a brain of any significant size works you're honestly looking at supercomputers the size of the planet with current technology.
Just throwing power at a problem like modelling the brain isn't going to make anything happen any time soon.
If we can accurately model an ant's brain, down to the individual neuron, in the next 100 years or so, I'll be impressed. If we can make an artificial algorithm of any kind that can surpass the learning ability and intelligence of your average kid up to being a teenager (so at least ten years of TRAINING them to do every task a human could do, and them handling it as well as a teenager could, from ZERO initial information), with the largest computers in the world for the next 100 years, I'll be impressed.
A brain is more than N CPU's executing X amount of cycles per second. Simulating that in any significant or useful way puts orders of magnitude on a task already out of range of the human race as a whole dedicating itself to ONLY doing that.
Consider this: You put a computer on the Internet with a "blank" memory. By ten years, even with *some* guidance, if it manages to learn to read and understand most Wikipedia articles just by learning what's come through it, that is INCREDIBLY impressive.
5x faster sounds ambitious, but not off the wall for 2015. It gives them 3 years during which time one might expect a 4x increase in speed. I expect that they will get the top spot for a bit before being passed by something else relatively quickly.
The top10 is quite interesting.
There's a bunch of PPC BlueGene, UltraSparc VIIIfx, Xeons, Opterons and GPUs.
As for the processors, this will be an interesting workout for them. I think like the UltraSparc VIIIfx, it will be hard for them to match the Xeon or Opteron for a while, but for supercomputing it is somewhat easier since brute FPU power works well which means you can tack on a huge vector FPU onto a relatively pedestrian CPU for excellent results.
The other thing of course is the interconnect. The VIIIfx got an extra really good boost by having a very good interconnect on die, giving it an exceptional rmax/rpeak efficiency.
I haven't seenmuch about a domestically developed killer interconnect yet, but they may very well have one which is good enough, or in this round they may go for something COTS.
The Chinese government has been pushing domestically produced processors for a while now, and they have been improving from a very los start, very quickly. I don't doubt that they will catch up in many areas. IPC on a general purpose CPU is a game of diminishing returns.
SJW n. One who posts facts.
It took a long time for evolution to put a person together. A machine will get to our level a bit faster. My phone does face recognition and tracking in an instant. That still amazes and worries me.
If we can accurately model an ant's brain, down to the individual neuron, in the next 100 years or so, I'll be impressed.
If you can do that without violating any patents, EVERYONE would be impressed.
And any answer it gives will be deemed 100% correct, even if it's not. Get used to, "2 + 2 = 5, and it always has. How dare you question the accuracy of our machine!!"
You seem to be assuming that modern jetliners flap their wings to fly. Somehow I'm not sure that the engineering approach to mimicking nature involves copying nature piece by piece. Therefore I'm not sure whether simulating anything to the degree you mention is anything else than a matter of idle curiosity. It may still turn out that we may not need that at all.
Ezekiel 23:20
The reason for the size is that the censor chips will have to watch what the working chips try to share with other working chips.
I wouldn't be surprised at all to discover that their domestic MIPS chips are basically just process shrunk Alphas. You can get some mileage out of just shrinking the transistors used, but I already shudder for the power bill this supercomputer is going to have.
I read the internet for the articles.
No, your phone does edge detection to get a series of numbers. There's a HUGE difference between that and actual intelligence.
I can program a computer in seconds to do maths that's beyond any human's capability to work out in their lifetime. That's not intelligence.
Now if you could say on each image that "this is me", "this isn't me", and it could build up a database of people that should and should not be authorised and OVER TIME learn on it's own without just having a bunch of statistics like "> 20% green = > 90% probability", then you'd have some mild form of intelligence. Otherwise you just have heuristics, which are 0.000001% of how actual intelligence operates.
What you're describing is a black box that gets an answer right in one scenario when fed with ALL available data. I have no data regarding what my father looks like if he dressed up in drag and shaved his head and I looked at him from the back - but I'd stand a pretty damn good chance of recognising him if he did that and my brain got that fresh, never-before-seen data.
Computers currently only form patterns that you inform them could exist in the data (in some way). They never form patterns of their own, but even pigeons are capable of that. Feed a pigeon at certain times of the day and it associates *SOMETHING* that happens at that time of day with feeding. Literally, it gets "superstitious" and does things like bang its head on a wall because that happened to coincide with feeding last time. It posits a theory, tests it repeatedly, takes that new data into account, and changes the theory as necessary. And that's a pigeon.
Computers, currently, do NOT do this, unless informed to do this, which is another matter entirely. Humans are not numbers-crunchers and computers are not hard-wired biological circuits joined by physical processes dependent on billions of interactions that change every second. Although either one can simulate the other, to some degree, it's difficult, long-winded, and like trying to play Half Life 3 on a Turing machine.
Phillips TVs had MIPS in them last time I checked (1-2 years ago, just before Phillips changed their name (flogged the division off to some international consortium, IIRC)).
Also FatPhil on SoylentNews, id 863
You don't carbon-copy things. That just makes a brain, or a virtual brain, not a computer-that's-as-good-as-a-brain, which is what we're discussing.
But the neuronal connections (10,000 per neuron, constantly changing) and time-slices that you'd need to get even close to simulating the number of interactions in an ant's brain when it smells a certain scent-trail? You might JUST about be able to approach that complexity if you turned the whole Earth into a computer with the nanoscale technology.
It's not a question of brute-force, or shrinking down, it's seriously an order-of-magnitude number of orders-of-magnitude out in terms of the complexity required to do something useful (and the criteria for that I'm setting deliberately low).
Sure, there are more efficient ways of doing things (hence computers manipulate numbers faster than anyone on earth even could), but this isn't about efficiency. This is about complexity, which is the opposite end of the scale. We just don't have enough complexity in any earth-based system built on technology big enough to put into chips and powered by electricity to do anything close to even a theoretically-more-efficient amount of intelligence.
The scale is immense. Of course, we can perform huge numbers of amazing tricks (face-recognition, etc.) but in terms of getting somewhere where we can replace a human with a computer on any significant task and not notice? Orders of magnitude away.
You're talking billions of computers with 10's of thousands of connections to each other than can be programmatically changed at will and operate at the speed-of-light to even be able to approach the equivalent of a brain of a small mammal, if you're lucky. It's a different kind of technology, yes, but it's the complexity of what we have in our heads and the complexity of "emulating" that via what we have in our computers that really blow the idea out of the water for centuries to come.
Hell, a couple of large A* searches can pull a modern computer to its knees with even the most efficient code. And, in terms of complexity, that's nothing and is optimised to their preferred method of data processing.
And please, although the Chinese government is very corrupt, it is not more corrupt than US government or US corporations.
Transparency International says otherwise. Way, way otherwise.
China is second world. Commies are second, democracies and those aligned are first, third is everyone else, it's political alignment, not social or economic.
Just bought a new quantum computer, but I'm uncertain how it works.
Because they're not going to let us do to them what they know they've done to us.
Hey free trade everybody.. get your free trade riiiight over here! Free movement of goods services and people ... hey ... whatsamatter with you??? Doncha' love FREEDOM????
We should be introducing out first Exaflop machines.
"To those who are overly cautious, everything is impossible. "
The point was technology (my phone) hasn't been in development for over a billion years. Tommy Flowers only put Colossus together in the 1930s. All you've said is true given less than 100 years of effort. I impressed that a tiny tiny phone does "edge detection". I stopped reading after, I can't help if you're not happy.
But the neuronal connections (10,000 per neuron, constantly changing) and time-slices that you'd need to get even close to simulating the number of interactions in an ant's brain when it smells a certain scent-trail? You might JUST about be able to approach that complexity if you turned the whole Earth into a computer with the nanoscale technology.
I'm sorry, but I'm not buying that. Are you suggesting that the difference of the density of features between brain tissue and anything that we can reasonably manufacture is on the scale of 10^10? (6*10^24 kg for Earth and 10^-6 kg for an ant brain, with negligible difference in density between the two, so 30 orders of magnitude of difference in volume) This is ridiculous, that would mean that we would have to simulate a single neuron (20 um?) using a 100km-sized device. That seems really ridiculous, no matter how complex a single neuron's interaction with the environment is.
Ezekiel 23:20
Like it or not, the Chinese have as much right to take what others have made and build on it as the US had for most of the last century. Remember, modern rocket technology was invented in Germany and the computer was invented in UK; and a lot of other things Americans think were their own inventions came from elsewhere.
your "assumptions" and "i will be impressed if..." are overly pessimistic. ever heard of http://bluebrain.epfl.ch/ ?
I wonder if the chinese will impose an export ban, preventing anyone "leaking" their technology to the USA?
politicians are like babies' nappies: they should both be changed regularly and for the same reasons
That's some serious content filtering they got going on. Complete with chastising AI.
Life is not for the lazy.
You are way off. 1st off the neuro nets in use today doing amazing things are based off a really really simplistic abstracted concept of how neurons work. You don't have to simulate the whole thing, you can crudely approximate it and get world changing results and we've already done this using a concept from the 1940s. With advances in our knowledge we can make the approximation more accurate without simulating all the details--- in fact, I worked with a guy for a short bit who was working on a mainframe that was simulating 100s of neurons down to the ion flow and chemistry and they were not doing it thinking we'll buy a bigger computer someday, they were doing it to discover more which would be used in an abstraction that didn't require all the complexity of the physical world and millions of years of attempts to find - it is much better for a digital machine to do digital work not to simulate analog which is what you are thinking of. We don't simulate analog on computers; we convert analog into digital then use digital techniques to simulate the analog - usually highly abstracted as well -- because an approximate result is enough (it is analog, there is no exact result you can only ever hope to approximate it.) Given the rate of high-resolution analog simulation of brains going on now it wouldn't take a planet sized computer to do it but it is out of practical reach for now.
They'll simulate a human brain scan SOON. it won't do much for quite a while afterwards - once it does anything meaningful we'll read the press release. The first ones will be simple neurons networked to model the brain scan; something that is probably possible today if somebody would fund it. Brain imagining should be far enough, like probably in the last few years. I don't think a cloned brain network will work because that network is heavily influenced by external factors (like behavior at operating temp) which is why people are trying to improve the abstract model and simplify it... it could be the 1st ones are still dead simple but they know how to adjust the scanned brain's network to compensate. Me, I think many of those difficult problems will be solved by AI powered tools. We you do get something that begins to approximate a mind I think the scifi people will be wrong, the machines will be erratic and insane. It'll take a long time before they are functional (even then likely by self-training simulations which would still produce insane AIs that are not predictable.) If you mess with AI programming a bit you'll get what I'm talking about. Realistically, the practical AI is what the effort is put into-- the people who think of AI as Applied Intelligence because that is mostly what is going on - we can grasp and understand narrow niches like spam filtering, plus it produces useful stuff now - the old AI stuff likely will never grasp it all, or if they do they can't know they actually achieved it. Not that such research is not beneficial long term.
Democracy Now! - uncensored, anti-establishment news
Comment removed based on user account deletion
Unclean hands aren't an excuse for further wrongs.
Right. No, your other right. No, the other other right.
As long as there are companies trying to reverse engineer the Chinese chips then if there is a backdoor, it will be found. If a backdoor is found it will be really damaging to their reputation.
It's one thing to be suspected of espionage, it is another thing altogether to be caught putting backdoors into their products.
Don't know something? Look it up. Still don't know? Then ask.
I don't think that by "brute force" he meant doing quantum state modeling.
FWIW, there is currently in progress an attemt to model the entire nervous system of C. elegans. Admittedly that's only 35 neurons or so, but once that is done we should know what parts of the neuron it's important to model.
So here we have the need for a neuron, or perhaps synapse, level model of the neural system as a "brute force" approach. I.e., we model the neural system as it exists, without understanding it.
OTOH, I've never made an estimate of how many megaflops that would require. If someone wants to claim that 1 exaflop would be enough, my response would be "Could be".
I think we've pushed this "anyone can grow up to be president" thing too far.
It doesn't seem to have damaged Microsoft very much.
I think we've pushed this "anyone can grow up to be president" thing too far.
Has anyone heard about a project to develop a 100 PF machine by a group in Santa Fe, New Mexico? There were rumors locally a couple of months ago, but I can't find any other information about it.
Before anyone's panties get in too tight a bundle here, we probably should actually follow the link and do that quaint, old-fashioned thing called "reading", which would lead us to notice that the person interviewed says: "I think in the future, as China tries to reach for exascale computing, the designs of these new supercomputers could fully rely on domestic processors. I wouldn't dismiss the possibility." and "But of course, we could take a different road, and end up using both foreign and domestic chips".
Rest easy, they don't even have their architecture figured out yet.
"Ahh! I see you're in that indeterminate Schrodinger state where - oh, uh
http://en.wikipedia.org/wiki/Loongson
http://en.wikipedia.org/wiki/Ingenic
> Now if you could say on each image that "this is me", "this isn't me", and it could build up a database of people that should and should not be authorised and OVER TIME learn on it's own without just having a bunch of statistics like "> 20% green = > 90% probability", then you'd have some mild form of intelligence. Otherwise you just have heuristics, which are 0.000001% of how actual intelligence operates.
Cute. From which orifice did you pull the 0.000001%? Do share with the rust of us your elaborate knowledge concerning the workings of intelligence. What is the other 99.999999%? How did you obtain this information?
Oh, give it a rest, will you?
This is very much the way things work in real life. If I see somebody doing something that I think I would like to do, of course I'll learn from it and, in your words "steal their idea".
And, at the same time, it looks like a lot of people on this list find it obviously right to make a 'backup copy' - or 10 or 100 - of any music or data CD, more or less. File sharing is commonly seen as acceptable; but if the Chinese, the devils, do something similar, then it is THEFT and deeply immoral. Right.
And the evidence of their "backdoor" is what again?
Don't know something? Look it up. Still don't know? Then ask.