Slashdot Mirror
US Regains Supercomputing Crown, Besting China and Japan
dcblogs writes "The U.S., once again, is home to the world's most powerful supercomputer after being knocked off the list by China two years ago and Japan last year. The top computer, an IBM system at the Department of Energy's Lawrence Livermore National Laboratory, is capable of 16.32 sustained petaflops, according to the Top 500 list, a global, twice a year ranking, released Monday. Despite the continuing strength of U.S. vendors globally, when China's supercomputer took the top position in June, 2010, it seemed to hit a national nerve. President Barack Obama mentioned China's top ranked supercomputer in two separate speeches, including his State of the Union address last year."
Each time I read a story like this I can't help think there are a bunch of faster machines that they don't tell us about.
At least in the *specific* performance characteristic of 64bit precision linear algebra, it's perfectly likely that the biggest player is reported.
In the cases where secrecy is probably preventing you from knowing about it, it probably is optimized for 32-bit precision floating point and/or large storage throughput to fuel data mining.
Of course, then there are collections of systems that could probably easily place in the list that are at least moderately well-known but not submitted, if it wouldn't be a financial catastrophe to take it down for a few days to dedicate to an xhpl run. An EC2 datacenter comes to mind.
XML is like violence. If it doesn't solve the problem, use more.
I'd rather have a big fraction of our workforce be highly competent in mathematics, than have a computer that's marginally faster than any other.
One wins a pointless pissing match, the other provides a much more solid basis for real strength and prosperity.
Besides, all this really shows is that China will lend us enough money for us to buy computer components built an assembled throughout the world.
Way to go !!
Moore wasn't an IBM employee, so Turek hasn't really heard that much about him.
Actually, the top 500 "competition/list" has been moderately useful for transitioning the world from big iron supercomputers sold by single corporations into the modern commodity/cluster (beowulf) model that costs far less and scales (as one can see) almost indefinitely large/fast for a certain class of linear or embarrassingly parallel problems. It is also the case that some of the problems that are solved using the larger of the computers built (which with the exception of corporate entries aren't really built "just for bragging rights") are both interesting and potentially of some value, either intellectual or monetary, to society. So it isn't, really, just a matter of bigger dicks (although the top ten does have a certain amount of that going on, where for decades some companies were conspicuous by their absence, and only got there eventually by basically building a machine with winning capacity and then giving it to somebody so that they could enter). Sometimes it is a matter of solving problems in nuclear physics or cosmology or cryptography or fluid dynamics that are NP complete or otherwise scale poorly enough that one is always hungry for cycles if one works in the field.
The question of whether or not the answers to those problems are worth the cost is a separate one, and by all means debate it, but be sure to do so in the context of all Big Instrumentation used in science. The LHC is a lot of money to -- maybe -- find the Higgs. Or not, again, to the tune of tens of billions of dollars. NASA routinely spends/spent tens of billions of dollars to lift humans and e.g. the Hubble into orbit -- the Hubble gives us enormous amounts of wonderful science but very little of that science is of direct (as opposed to indirect) benefit to (say) automobile mechanics, lawyers, owners of restaurants, farmers. The cost of a top 500 machine is in comparison cheap, and in some cases may even work on problems with a measurable expectation value that trickles back to the society that ultimately pays for it (outside of the noble cause of supporting the education and research system that has created a truly enormous amount of very concrete wealth by providing work for otherwise unemployed physicists and computer scientists and mathematicians and funding for the many science and math departments that trained them and whose faculty participate). Personally, I think it is well worth it, but I've spent a good fraction of my life attached to that particular teat (although I'm not, currently) and don't pretend to be completely objective here.
I am, OTOH, pretty well informed about cluster computing, while having absolutely no dog in the top 500 race.
rgb
Even when the experts all agree, they may well be mistaken. --- Bertrand Russell.
It's too bad these computers are in the hands of governments instead of people who could do something productive with them other than delegating processes to espionage, war, etc. Even if the computers did major calculations that could point to policy change in government, there are still politicians with motives and bureaucrats within agencies that have to protect their jobs...and hence would never use it productively. So, what's the point?
The fastest computer in the world is useless if it's not being used for a productive purpose.
42
Just a little interesting thing I noticed while reading up on this.
Most American media refers to it as an "American" supercomputer first and foremost, and an "IBM" supercomputer second.
Most non-American media refers to it as an "IBM" supercomputer first, and an "American" supercomputer second.
Not really wrong either way - it's a big win for both the US and for IBM - but it's interesting to see the little differences.
Seriously, if they are migrating to new supercomputers, what, each year or two, what are they doing with the "scraps"? Do they refurbish them to be used elsewhere? do they sell them as "used goods" to other companies?
I can imagine taking quite a bit of time to install/uninstall such installations, so if its going to stay 2 years and it takes 3 months to setup/remove, are the installation/uninstallation crews constantly on site? or maybe it takes much less time to do so?
does anyone has any idea on this? I mean, that's a lot of stuff to "just" throw away when upgrading. Something useful must be done with the "obsolete" equipment, no?
Slashdot. Unreadable news to annoy nerds. - wonkey_monkey
in 6 months another country will take the crown
RISC and CISC are no longer really meaningful distinctions. The new POWER systems have a lot of non-RISCy SIMD extensions and such, as do the SPARCs used in the K supercomputer. Likewise, the modern CISC processors (read: x86) have become more RISC-like, at least internally.
I actually think the modern x86 way is a really good compromise (although the x86 implementation of the idea ranges from terrible to tolerable). Internally, use a simple, efficient RISC-like micro-op processor, while presenting a CISC-like interface to the rest of the system. You get most of the advantages of both (the speed of RISC, the instruction density of CISC) as well as some unique ones (as long as you keep the CISC layer the same, you can change the internals all you want while keeping full software compatibility). The only significant downside is a small loss of die space for the translator (usually only a few percent of the die (half of which is usually a massive L3 cache anyways)).
Now if only that CISC layer wasn't the mess known as x86-64...
You commit several severe logical errors during this (to say nothing of your grammar). Let's look at it in depth.
Your first fallacy is assuming all cores are equal. Notably, that PowerPC A2 is running at half, maybe even a third, the clock speed of the i7. That gives it much better efficiency, yes - but it also gives much higher latency. In a supercomputer processor, that's no problem - throughput is the only important factor. But for a desktop? Latency matters.
Had you read more, you would also know that one of those 18 cores is reserved for the OS and interrupt handling, and does not contribute to actual programs. Another core is disabled for yield purposes - they expect one core to be defective, and plan for it.
You're also comparing unlike processors. The A2 targets efficiency and parallel throughput - it does not need to deal with "what if the program being run is only single-threaded, and we can't throw cores at it?". Meanwhile, the i7 has to consider what a "regular user's" workload is, and how it can best accommodate it. With tricks like turbo boost, or even the various sleep modes - the i7 only draws that much power under full load; under light use, it draws far less; the A2 does not have to worry about sleep-mode power draw or low-load power consumption, as it's intended to run at full-throttle at all times.
May I also ask which specific i7 you were referencing? 22nm implies it's an Ivy Bridge series, but none that have yet been released are 8-core processors, nor do any of those use 135W. Some of the slightly-older Sandy Bridge-E series have specs more similar to what you specified (8 cores, but only 130W), but those are built at 32nm and are not exactly "latest", although they remain "greatest".
At least for desktop processors, and there is your next error. The i7 is a desktop/workstation processor. Costs about $300 for a low-end one, and maybe $1000 for a top-of-the-line my-dick-is-bigger-than-yours overcompensator. I cannot find pricing on the A2, but I would easily bet that it's closer to $3000 than $300. It's a server chip - if you wanted a fair comparison, you would have compared it to a Xeon E7. Which, as an aside, seems due for a refresh - the newest E7 was released over a year ago, and was based on the truly-outdated Westmere architecture.
You go on to argue *against* yourself. If 18,000 Intel processors gets you 2.9 petaflops, that comes out to about 6200 processors/petaflop. The AMD system runs at an absurd 132,000 processors/petaflop, and the IBM 4,300 processors/petaflop. To me, that looks more like "Intel and IBM are neck-and-neck" than "IBM is far and away the best". Especially as we've not seen Intel's response, although a newer story on Slashdot is about Intel's MIC project - fifty cores per processor, and power consumption in kilowatts, not megawatts.
You then make the grave error of comparing products that have not even been announced yet (we've no guarantee an Xbox 720 or a Playstation 4 will even *exist*, much less use PowerPC chips).
Oh, and now you go beyond comparing server chips to desktop chips. Now you're treating game console chips as server chips as well. Which, granted, is true for the PS3's Cell (it *must* have been some non-technical marketing droid who decided on that, because that was one of the dumbest technical moves they could have made). But the Wii? A deliberately-pathetic, low-power chip chosen mostly for backwards compatibility.
Let's also look at the *other* consoles. The ones you *didn't* reference. The PS2? A heavily-modded MIPS. The Dreamcast? A Super-H. The original Xbox? Why, our old friend Mr. Intel x86! And yet those were in the heyday of the PowerPC, when it seemed a genuine competitor on the desktop and server front.
Oh, and one more thing: the Mac. Apple quite famously switched *away* from PowerPC. And say what you will, but Apple is not an idiot. They could see that, at the time, IBM was going nowhere with it. Too much power usage for laptops and smaller desktops, and not enough computational po