Slashdot Mirror

← Back to Stories (view on slashdot.org)

Japan Eyes World's Fastest-Known Supercomputer, To Spend Over $150M On It (reuters.com)

Posted by msmash on from the need-for-speed dept.

Japan plans to build the world's fastest-known supercomputer in a bid to arm the country's manufacturers with a platform for research that could help them develop and improve driverless cars, robotics and medical diagnostics. From a Reuters report: The Ministry of Economy, Trade and Industry will spend 19.5 billion yen ($173 million) on the previously unreported project, a budget breakdown shows, as part of a government policy to get back Japan's mojo in the world of technology. The country has lost its edge in many electronic fields amid intensifying competition from South Korea and China, home to the world's current best-performing machine. In a move that is expected to vault Japan to the top of the supercomputing heap, its engineers will be tasked with building a machine that can make 130 quadrillion calculations per second -- or 130 petaflops in scientific parlance -- as early as next year, sources involved in the project told Reuters. At that speed, Japan's computer would be ahead of China's Sunway Taihulight that is capable of 93 petaflops. "As far as we know, there is nothing out there that is as fast," said Satoshi Sekiguchi, a director general at Japan's âZNational Institute of Advanced Industrial Science and Technology, where the computer will be built.

35 comments

  1. Question by Anonymous Coward · · Score: 1

    Does it meet the minimum system requirements for VR?

    1. Re:Question by darkain · · Score: 1

      Is this the new "flash without lag" meme?

  2. kill yourself by Anonymous Coward · · Score: 0

    objective solutions is now the name of the office that is holding child porn. they still finding a retard to work here in porto alegre.

  3. What do you do with the old ones? by HalAtWork · · Score: 1

    So what do you do with all the old supercomputers when they're too big/power hungry vs performance? Helluva paperweight.

    --
    Twinstiq, game news
    1. Re:What do you do with the old ones? by nairnr · · Score: 1

      You can upgrade them. It isn't like these are monolithic computers. You can upgrade network capabilites, change processors.

    2. Re:What do you do with the old ones? by aaarrrgggh · · Score: 1

      Like all things, there is a smashing point where improved energy efficiency pays the cost of an upgrade if performance needs are flat. When performance needs exceed the incremental upgrade options, you either settle for longer processing time, simplified process, or pay for a faster system.

    3. Re:What do you do with the old ones? by Waffle+Iron · · Score: 1

      So what do you do with all the old supercomputers when they're too big/power hungry vs performance? Helluva paperweight.

      If they're anything like me, they'll put it on a shelf in their basement.

      25 years later, they'll try turning it on again just for fun. One of four things might happen, with roughly equal probability:

      • 1. It works fine
      • 2. It boots, but acts really flaky with strange characters on the display
      • 3. It's a brick
      • 4. Some component goes up in a puff of smoke; goto 3
  4. Slitty by Anonymous Coward · · Score: 0

    Japan Slitty Eyes World's Fastest-Known Supercomputer, To Spend Over $150M On It

    - ft4U

  5. Post-patent supercomputing by Shane_Optima · · Score: 2

    It's strange how little people talk about the fact that we're approaching the end of this little plateau in the first silicon revolution. I'm sure this supercomputer is fast and all, but compared to what's economically feasible once the barriers are removed?

    The second silicon revolution will be marked by price freefall as soon as enough patents expire and enough high-output[1] factories can come on-line. Eventually (at the most, perhaps a few decades from now), a major world government will realize that if they buy their own factories, keep 'em cranking out single board machines and flash memory at full speed 24/7 and if need be use some superfluous power plant that was about to be decommissioned... they could build a supercomputer the likes of which the world has never seen. Maybe someone more knowledgeable than myself could do some back of the napkin estimates here about what should be possible?

    As an interesting aside, the power supply, display and input devices will end up becoming the most expensive parts of most consumer electronics, but I think the more interesting question is what the hell are they going to do with all of that computing power once the price floors give way? Protein folding, cryptography... and general AI.


    1. The "high output" bit being the kicker. I know very little of the details of chip lithography, so maybe there are hangups here I'm unaware of.

    1. Re: Post-patent supercomputing by johnsmithperson123 · · Score: 2

      The problem is that microchip foundries and dies are massive investments. Still, for a major world government, sinking $10 billion into a foundry wouldn't be an issue, especially since silicon looks to be bottoming out.

    2. Re: Post-patent supercomputing by Anonymous Coward · · Score: 0

      Building a super-computer is not a matter of adding more and more processing chips. There's a lot of work to get all the pieces working together.

      Besides, if any government wanted to do it, they could already. The Russians have a long history of it from the 1980s.

    3. Re: Post-patent supercomputing by Shane_Optima · · Score: 1

      The problem is that microchip foundries and dies are massive investments. Still, for a major world government, sinking $10 billion into a foundry wouldn't be an issue, especially since silicon looks to be bottoming out.

      Yes, but it's a problem that will *eventually* go away. The per-unit costs are just way too small, and the potential upside is way too large for it to not happen in our lifetimes[1]. A $10 billion cost is on the order of the LHC, but unlike the LHC such a project would offer massive ongoing practical benefits in addition to theoretical research potential.


      1. [waiting for the "I'm 73, you insensitive clod!" remark]

    4. Re:Post-patent supercomputing by K.+S.+Kyosuke · · Score: 1

      I had similar ideas the other day. The real problem here is that you eventually hit a power wall as you continue to deploy new stuff en masse. Nevertheless, at least certain types of computing could constitute flexible demand, so you might want to power them with any extra generation you have at the moment.

      --
      Ezekiel 23:20
    5. Re:Post-patent supercomputing by Shane_Optima · · Score: 1

      I had similar ideas the other day. The real problem here is that you eventually hit a power wall as you continue to deploy new stuff en masse.

      That's why I explicitly mentioned superfluous power plants. There was just a story the other day about coal plants closing. Well... keep one online. How many flops can a single, large coal plant going full-tilt give you? I bet it's a lot. I bet the bottleneck, from a super-project perspective (not necessarily home user), isn't going to be the electricity.

    6. Re:Post-patent supercomputing by Shane_Optima · · Score: 1

      I had similar ideas the other day. The real problem here is that you eventually hit a power wall as you continue to deploy new stuff en masse. Nevertheless, at least certain types of computing could constitute flexible demand, so you might want to power them with any extra generation you have at the moment.

      It's also worth noting this doesn't prevent an explosion of dirt-cheap flash memory from happening at some point which has its own set of interesting implications. Particularly if high quality video sensors also fall in price...

    7. Re:Post-patent supercomputing by K.+S.+Kyosuke · · Score: 1

      I thought you meant the manufacturing part. I had the operations more in mind. What I thought of was something similar to DESERTEC, only without long-distance power transmission and with on-site computation instead. Just screw a computing module on the back of each solar panel and then fiber it up. You could power stuff with coal from a central location but that way, you'd never get rid of the dirty stuff.

      --
      Ezekiel 23:20
    8. Re: Post-patent supercomputing by Anonymous Coward · · Score: 0

      Yea you just dong go setting up a fab shop and start making large amounts of good chips like it's potatoes

    9. Re:Post-patent supercomputing by 110010001000 · · Score: 1

      You don't achieve AI by just throwing processor power at it. Cost isn't a concern here. If it cost $400 billion to achieve AI someone would spend that.

    10. Re:Post-patent supercomputing by Shane_Optima · · Score: 1
      Given that AI, particularly the sort of radically game-changing stuff we're likely to care about, is almost certainly going to be formed through self-modification and testing, it's hard to see how massively increased CPU power isn't going to make this easier. Also, I mentioned flash memory for a reason; storage should fall in price along with processing power. Being able to predict what humans do or say next (in text, audio or video) seems crucial here. This isn't about Deep Blue so much as Watson. Neither is recognizable as what we mean by "real" AI but the latter obviously appears to be headed in the right direction... for example, when it got a question wrong (by misunderstanding the sort of answer the category was looking for), it paid attention to the answers the human contestants were giving and quickly adapted so that it could give the correct answers for that category as well. And, as I recall, it accomplished that in part through massive trial and error and examination of large of text it had ripped from the internet. Being able to predict human behavior (first developing useful abstractions used to quantify said behavior) seems like a doable goal in principle and is a plausible path to general AI, but it requires a hell of a lot of processing power, particularly if we're talking about analyzing/predicting emotions from audio and video files instead of merely text.

      If you doubt the importances of this approach, note that our own intelligences were bootstrapped from an existing animal intelligence built around 3d navigation and instinctual emotional recognition.

      If it cost $400 billion to achieve AI someone would spend that.

      "If computing was in any way useful, Congress would have given Babbage millions of dollars to build his Analytical Engine." (But they didn't. Instead, it was over a hundred years before governments got serious about computing.)

      There are other significant advantages in having a supercomputer orders of magnitude more powerful than what we have today, even without general AI. Not every problem is massively parallelizeable, but many are.

  6. For what purpose? by fahrbot-bot · · Score: 1

    Probably just want to explore all of No Man's Sky ...

    --
    It must have been something you assimilated. . . .
    1. Re:For what purpose? by Anonymous Coward · · Score: 0

      Obviously, it's to determine the question for which the answer is 42.

      What more does anyone need?

  7. Japan eyes? by Anonymous Coward · · Score: 0

    Racist! -_-

    1. Re:Japan Eyes? by ClickOnThis · · Score: 1

      Is that the name of the new supercomputer?

      If not, learn to write proper titles.

      I think you need to learn to recognize the difference between a noun and a verb.

      --
      If it weren't for deadlines, nothing would be late.
  8. Sound impressive! by scunc · · Score: 1

    But can it play Doom?
    ---
    Asking the important questions since 2016

    1. Re:Sound impressive! by Anonymous Coward · · Score: 0

      Sad that the highest modded remark on this story is at 2, and that the most technical story in ages on here has the fewest comments. RIP Slashdot.

  9. manishs == crap by Hognoxious · · Score: 1

    WTF is ÃZNational?

    --
    Confucius say, "Find worm in apple - bad. Find half a worm - worse."
    1. Re:manishs == crap by Anonymous Coward · · Score: 0

      Asian national.

    2. Re:manishs == crap by Anonymous Coward · · Score: 0

      Slashdot is taking the Unicode boycott to the next level. Now they insert mojibake where the original doesn't even have a non-ASCII character.

  10. $150 million? by ClickOnThis · · Score: 1

    That's hardly pocket-change, but it seems cheap for what would be the world's fastest-known supercomputer.

    For comparison, Tianhe-2 (in number 2 spot) cost about $390M to build, and Sunway TaihuLight, the current number 1, went live in June of this year at a cost of $273M.

    --
    If it weren't for deadlines, nothing would be late.
  11. Japan Eyes? by Anonymous Coward · · Score: 0

    Is that the name of the new supercomputer?

    If not, learn to write proper titles.

  12. Probably x86 clusters by rsilvergun · · Score: 1

    which are about to get a whole lot cheaper next year when AMD launches their Zen CPUs. Assuming AMD isn't just lying about price/performance (possible) for the first time in 10 years they're going to be competitive with Intel, even at clocks per watt. That'll have a huge impact on the cost of CPUs. Already leaks indicate they'll be pricing at about 1/2 Intel's i7 for an equivalent CPU.

    --
    Hi! I make Firefox Plug-ins. Check 'em out @ https://addons.mozilla.org/en-US/firefox/addon/youtube-mp3-podcaster/
    1. Re:Probably x86 clusters by Anonymous Coward · · Score: 0

      I've not heard they will be able to match the i7. Tech speculation is that they may be competitive with i3 and i5.

      Look, if we get a dramatic leap forward in CPU development, and the first solid competition in the x86/x64 market since about 2006, we all win. However it is quite a stretch to believe that AMD goes directly from being on life support and a CPU also-ran, to "...pricing at about 1/2 Intel's i7 for an equivalent CPU". And by implication, the technology leader in this space.

      Also, the architectural information I've seen suggests that AMD is playing with the same basic designs as Intel. There are no 'unbelievable design breakthroughs' there. No vector processing, no Cell-like technology, no VLIW capability, no FPGA hybrids, no SIMD designs. No neural networks, no AI.

      What AMD has, is a solid APU design. They have HBM (which has no obvious application to the CPU). They have a solid history of tech leadership from 2003 - 2006. And of course they are still churning out good discrete GPUs. However the Bulldozer architecture failed them big-time and now they are riding the edge of disaster. Zen has to be a success now, since AMD has no financial cushion left anymore.

  13. Black Friday by Anonymous Coward · · Score: 0

    So what's the Black Friday discount on this? I'll take two.

  14. US commissioned two exaflop computers by peter303 · · Score: 1

    Paying nine figures hoping for 2019 or so. With mulriple countries chasing the exaflop, so good ideas may come out of it.

  15. Core speed not as important for HPC by Anonymous Coward · · Score: 0

    For most of the problem sets that HPC/Supercomputers deal with, the speed of an individual core is not as important as bandwidth and latency between individual cores. All most all of these problem sets can only, practically, be approached with massive parallelism. The newer Phi chips from Intel top out at 1.5Ghz but there up to 72 cores on a single chip and they're communicating with each other over a 400GB/s mesh. The fastest Xeon processors top out at 3.5ghz in a 4 core package, so you would need ~31 to compete with one Phi on a Mhz basis. You can put 16 cores into one box but QPI (inter-processor bus) tops out at ~38GB/s, 10% the same bandwidth. You'd also need another server for the other 15 cores and that would need to go over Ethernet or Infiniband, which in theory can do 40GB/s but are practically limited to 10-20GB/s because of the PCI-Express bus on most servers. On top of this, you have increased latency with both QPI and networking. So yes, you have faster nodes but they are spending a lot more time stalling out when communicating with other nodes. You also have the added hassle of requiring more racked hardware and the supporting network infrastructure.
     
    With that said, I don't know where you're getting your information that AMD has no VLIW or SIMD. Any AMD APU (Radeon on chip) supports VLIW, they have all shipped with SSE (SIMD support) since at least 2011 and AVX support since 2012. They do lag behind a bit, AVX2 was only introduced on the Carrizo last year, but they most certainly have it. I haven't developed against them so I can't speak for the implementation speed though.
     
    Finally, the push about neural networks, swarm computing and AI are a great deal of marketing spin. Nothing Intel or nVidia ships today has a magic "AI" instruction set. Massive parallelism, SIMD, AVX, etc all make this type of computing much easier but at the end of the day you're still going to have to code up the "AI" for your problem set/domain.