看
到中国安装了世界最快的超级计算机，美国政府机关和企业立即行动起来，要恢复美国在这项技术中的领导地位。在气候研究、产品设计和武器开发等领域，超级计算机都是一个至关重要的工具。
Reuters
中国研制的天河一号超级计算机
这次的挑战可能要复杂得多。要实现运算性能的下一步重大赶超，就需拥有多至10亿台电脑组成的系统，并在编程方面取得突破，从而让这些电脑发挥作用。而国会共和党人下定决心要减少赤字，说服他们为这类开发活动提供补贴恐怕是很难的。
加州劳伦斯•伯克利国家实验室(Lawrence Berkeley National Laboratory)是超级计算机的重要用户，实验室副主任西蒙(Horst Simon)承认赶超不容易。但他说，可以认为，如果没有新一代超级计算机系统，一些重要的科学问题将得不到解决。他说，这其实是一个经济竞争力问题、一个国家安全问题。
超级计算机有房间那么大，由数十万个处理单元组成，可用于同时解决多个涉及数字密集运算的问题。这种系统起先是用于与国防有关的繁重工作，如破译通讯密码等，但现在已经普遍使用于新车或药物设计等任务。
中国新一代超级计算机名为“天河一号A”，位于天津，上个月首次对外公布。本周美国新奥尔良的一场超级计算机大会上，“天河一号A”预计会成为热门话题。大会公布的新排行榜将会确认初步估计，证明“天河一号A”在标准运算测试中的得分高于此前为世界最快的“美洲虎”(Jaguar)系统。“美洲虎”位于美国田纳西州橡树岭国家实验室(Oak Ridge National Laboratory)。
但美国超级计算机专家说，给美国带来挑战的不仅是“天河一号A”。世界上性能第三的超级计算机系统也属于中国，而在上述大会半年发布一次的超级计算机500强排行榜中，中国目前拥有42台，在大型超级计算机的安装数量方面仅次于美国。
包括“天河一号A”在内的几台计算机，运算性能突破了一个“petaflop”（千万亿级，即每秒科学计算次数达到10的15次方）。中国有关机构已誓言五后至少有一套系统的性能达到50到100个“petaflop”。他们还希望在2016年到2020年实现一到10个“exaflop”（每秒10的18次方）的性能。
美国能源部设立的实验室──它利用超级计算机模拟大量核武器爆炸──正打算建造更加大型的机器，同时帮助起草一份更加雄心勃勃的“百亿亿次”计划。橡树岭国家实验室负责科学技术的副主任撒迦利亚(Thomas Zacharia)估计这些工作将在今后八年内额外需要40亿至50亿美元的联邦资金，以开发必需的硬件和软件技术。
从理论上说，这类系统可以通过将现有大型系统中使用的更多元件组合在一起来实现。但美国能源部的官员估计这样一台超级计算机将需要十亿瓦特的动力才能运转，大概相当于一座核电站所产生的动力。对“百亿亿次”计划而言，挑战在于要提高微处理器和其他芯片的性能同时大大减低它们的能量消耗。
如今的超级计算机大多使用的是英特尔公司(Intel Corp.)和高级微设备公司(Advanced Micro Devices Inc.)在个人电脑中广泛使用的微处理器的变体。中国最大的两个系统在这类芯片上增加了一种构件──图像处理单元，这类芯片可以使电子游戏中产生图像。
这种加速芯片有望在百亿亿次系统中发挥作用，目前它已经被一些制造商加以利用。硅图国际公司(Silicon Graphics International Corp.)周一计划推出一种名为Prism XL的超小型系统，它提供了一系列加速芯片可供选择──高级微设备公司或是英伟达(Nvidia Corp.)的图像处理单元，或者是初创企业Tilera Corp.的能加速各类运算的芯片。
英特尔正在研发基于其x86微处理器设计的加速芯片，它说这款加速芯片运行起来将比其他加速芯片更加便捷。德克萨斯州奥斯丁的初创企业 Smooth-Stone本周将更名为Calxeda，它计划向市场推出采用了ARM Holdings PLC微处理器技术变体的超级计算机和其他服务器，ARM Holdings是一家以手机为主营业务的公司。
中国工程师为天河一号A系统研制了本土的通讯电路，还研制出基于Sun 微系统公司(Sun Microsystems Inc.) Sparc技术的芯片。他们计划研制出完全自主创新的微处理器，在今后的机器中起到核心计算动力的作用。
这一设想实现后，中国对美国公司援助的依赖程度将大大降低，对美国政府出口限制的抵御能力也将增强，美国政府为限制接触到军事应用技术而设置了此类限制。另一家参与百亿亿次计划的美国机构国家科学基金会(National Science Foundation)的首席技术长穆诺兹(Jose Munoz)说，这些是美国需要关心的问题。
Don Clark
（更新完成）
（本文版权归道琼斯公司所有，未经许可不得翻译或转载。）
China's installation of the world's fastest supercomputer is galvanizing efforts by U.S. government agencies and companies to restore American leadership in the technology, a key tool in such fields as climate research, product design and weapons development.
Participants hope to outrace Chinese engineers in bringing a thousand-fold acceleration of today's most powerful machines -- replaying a crusade in the past decade that leapfrogged a supercomputer in Japan that briefly held the world speed crown.
This time, the challenges could be much tougher. Achieving the next major leap in computing performance could require systems with as many as a billion electronic brains, as well as programming breakthroughs to exploit them. And Republicans in Congress bent on reducing deficits may be hard to persuade to subsidize such developments.
It's not going to be easy,' concedes Horst Simon, deputy laboratory director at Lawrence Berkeley National Laboratory in California, a major supercomputer user. But he says a case can be made that important scientific problems won't be solved without a new generation of systems. 'It is really an economic-competitiveness issue and a national-security issue,' he says.
Supercomputers are room-sized collections of hundreds or thousands of processing units that can be used simultaneously on big number-crunching problems. Such systems were first used for defense-related chores, such as cracking communications codes, but now are widely used for jobs such as designing new cars and drugs.
The new Tianhe-1A machine in Tiajin, China, first publicly described last month, is expected to be a hot topic at a supercomputer conference this week in New Orleans. A new ranking to be released at the event will confirm preliminary estimates that the machine scored higher on standard computing tests than the Jaguar system at Oak Ridge National Laboratory in Tennessee, which previously had ranked as the world's fastest.
But the Tianhe-1A system is just one problem, U.S. supercomputer experts say. China, which also operates the third most powerful system, now accounts for 42 supercomputers on the semiannual Top 500 ranking, placing second behind the U.S. in large supercomputer installations.
Tianhe-1A among a handful of machines whose capabilities exceed a 'petaflop,' or a quadrillion scientific calculations per second. Chinese agencies have vowed to follow up with at least one system in five years that will reach 50 to 100 petaflops. Between 2016 and 2020 they expect to achieve one to 10 'exaflops,' or a quintillion operations per second.
Labs run by the U.S. Department of Energy -- which use supercomputers to simulate detonations of stockpiled nuclear weapons -- are planning to build larger machines while helping draft a more ambitious 'exascale' plan. Thomas Zacharia, deputy for science and technology at Oak Ridge National Laboratory, estimates the effort will take $4 billion to $5 billion in added federal funding over eight years to develop needed hardware and software technologies.
In theory, such systems could be constructed by stringing together more of the components used in today's large systems. But DOE officials estimate that such a supercomputer would require a gigawatt of power to operate, or about the output of nuclear power plant. A challenge for the exascale effort is raising the capacity of microprocessors and other chips while sharply reducing their power consumption.
Most of today's supercomputers use variants of the microprocessors that Intel Corp. and Advanced Micro Devices Inc. popularized in PCs. The two biggest Chinese systems augment such chips with an added building block -- graphics processing units, the variety of chips that render images in videogames.
Such accelerator chips are expected to play a role in exascale systems and are being used by some manufacturers already. On Monday, Silicon Graphics International Corp. plans to introduce an ultracompact system called Prism XL that offers a choice of accelerators -- either GPUs from AMD or Nvidia Corp., or chips from start-up Tilera Corp. that accelerate different types of calculations.
Intel is developing accelerator chips based on its x86 microprocessor design, which it says will be easier to program than alternatives. Smooth-Stone, an Austin, Texas, a start-up that is changing its name this week to Calxeda, plans to attack the market for supercomputers and other servers with a variant of the microprocessor technology from ARM Holdings PLC that is a mainstay in cellphones.
Chinese engineers developed homegrown communication circuitry for the Tianhe-1A system, and chips based on the Sparc technology that originated with Sun Microsystems Inc. They are planning to develop entirely original microprocessors to serve as the core calculating engines for future machines.
When that happens, China will be much less reliant on assistance from U.S. companies -- and less vulnerable to U.S.-government export controls designed to restrict access to technology with military applications. 'Those are things the U.S. needs to be concerned about,' says Jose Munoz, chief technology officer at the National Science Foundation, another U.S. agency involved in the exascale effort.
Don Clark
Participants hope to outrace Chinese engineers in bringing a thousand-fold acceleration of today's most powerful machines -- replaying a crusade in the past decade that leapfrogged a supercomputer in Japan that briefly held the world speed crown.
This time, the challenges could be much tougher. Achieving the next major leap in computing performance could require systems with as many as a billion electronic brains, as well as programming breakthroughs to exploit them. And Republicans in Congress bent on reducing deficits may be hard to persuade to subsidize such developments.
It's not going to be easy,' concedes Horst Simon, deputy laboratory director at Lawrence Berkeley National Laboratory in California, a major supercomputer user. But he says a case can be made that important scientific problems won't be solved without a new generation of systems. 'It is really an economic-competitiveness issue and a national-security issue,' he says.
Supercomputers are room-sized collections of hundreds or thousands of processing units that can be used simultaneously on big number-crunching problems. Such systems were first used for defense-related chores, such as cracking communications codes, but now are widely used for jobs such as designing new cars and drugs.
The new Tianhe-1A machine in Tiajin, China, first publicly described last month, is expected to be a hot topic at a supercomputer conference this week in New Orleans. A new ranking to be released at the event will confirm preliminary estimates that the machine scored higher on standard computing tests than the Jaguar system at Oak Ridge National Laboratory in Tennessee, which previously had ranked as the world's fastest.
But the Tianhe-1A system is just one problem, U.S. supercomputer experts say. China, which also operates the third most powerful system, now accounts for 42 supercomputers on the semiannual Top 500 ranking, placing second behind the U.S. in large supercomputer installations.
Tianhe-1A among a handful of machines whose capabilities exceed a 'petaflop,' or a quadrillion scientific calculations per second. Chinese agencies have vowed to follow up with at least one system in five years that will reach 50 to 100 petaflops. Between 2016 and 2020 they expect to achieve one to 10 'exaflops,' or a quintillion operations per second.
Labs run by the U.S. Department of Energy -- which use supercomputers to simulate detonations of stockpiled nuclear weapons -- are planning to build larger machines while helping draft a more ambitious 'exascale' plan. Thomas Zacharia, deputy for science and technology at Oak Ridge National Laboratory, estimates the effort will take $4 billion to $5 billion in added federal funding over eight years to develop needed hardware and software technologies.
In theory, such systems could be constructed by stringing together more of the components used in today's large systems. But DOE officials estimate that such a supercomputer would require a gigawatt of power to operate, or about the output of nuclear power plant. A challenge for the exascale effort is raising the capacity of microprocessors and other chips while sharply reducing their power consumption.
Most of today's supercomputers use variants of the microprocessors that Intel Corp. and Advanced Micro Devices Inc. popularized in PCs. The two biggest Chinese systems augment such chips with an added building block -- graphics processing units, the variety of chips that render images in videogames.
Such accelerator chips are expected to play a role in exascale systems and are being used by some manufacturers already. On Monday, Silicon Graphics International Corp. plans to introduce an ultracompact system called Prism XL that offers a choice of accelerators -- either GPUs from AMD or Nvidia Corp., or chips from start-up Tilera Corp. that accelerate different types of calculations.
Intel is developing accelerator chips based on its x86 microprocessor design, which it says will be easier to program than alternatives. Smooth-Stone, an Austin, Texas, a start-up that is changing its name this week to Calxeda, plans to attack the market for supercomputers and other servers with a variant of the microprocessor technology from ARM Holdings PLC that is a mainstay in cellphones.
Chinese engineers developed homegrown communication circuitry for the Tianhe-1A system, and chips based on the Sparc technology that originated with Sun Microsystems Inc. They are planning to develop entirely original microprocessors to serve as the core calculating engines for future machines.
When that happens, China will be much less reliant on assistance from U.S. companies -- and less vulnerable to U.S.-government export controls designed to restrict access to technology with military applications. 'Those are things the U.S. needs to be concerned about,' says Jose Munoz, chief technology officer at the National Science Foundation, another U.S. agency involved in the exascale effort.
Don Clark
没有评论:
发表评论