【终极转载】美国计算机历史博物馆展览-超级计算机之3对速度的需求
Backplane wiring of the CDC 7600 supercomputer
The wiring was done by hand and had to be precise to within a fraction of an inch to optimize the speed of the machine.
有道翻译
CDC 7600超级计算机的背板接线
接线是手工完成的，必须精确到几分之一英寸才能优化机器的速度。
Various Paths to Supercomputing
The earliest computers were sequential machines, processing one number at a time. Some supercomputers increased processing speed by working on many numbers at once.
Vector supercomputers like the Cray-1 perform calculations on all the numbers in a sequence by filling a pipeline with them and processing them assembly-line style.
Parallel computers combine multiple processors into a larger system. These spread a complex calculation among many processors working independently, and then reassemble the results into a single answer. In SIMD computers (“single instruction, multiple data”), all processors execute the same instructions at the same time on different data. In MIMD computers (“multiple instruction, multiple data”), the processors needn’t operate in lock-step, but still exchange results.
Clusters the most common type of supercomputer built today — are groups of computers connected by networks and coordinated by software. Large clusters can contain a million linked computers.
有道翻译
超级计算的各种途径
最早的计算机是顺序机器，一次处理一个数字。一些超级计算机通过同时处理许多数字来提高处理速度。
像Cray-1这样的矢量超级计算机通过用它们填充管道并处理它们的装配线样式来对序列中的所有数字执行计算。
并行计算机将多个处理器组合成更大的系统。这些在独立工作的许多处理器之间进行了复杂的计算，然后将结果重新组合成单个答案。在SIMD计算机（“单指令，多数据”）中，所有处理器在不同数据上同时执行相同的指令。在MIMD计算机（“多指令，多数据”）中，处理器不需要以锁步方式操作，但仍然交换结果。
集群当今构建的最常见类型的超级计算机 - 是通过网络连接并由软件协调的计算机组。大型群集可以包含一百万个链接的计算机
Cray-1 diagram
The Cray-1 featured a very large memory as well as both sector and scalar functional units.
有道翻译
Cray-1图
Cray-1具有非常大的内存以及扇区和标量功能单元。
Gordon Bell: Three Phases of Supercomputing
有道翻译
戈登贝尔：超级计算的三个阶段
The Need for Speed
Speed isn’t just a luxury for the impatient. Supercomputers do what couldn’t otherwise be done, by simulating systems that are dangerous, expensive or impossible to try in real life. This provides a “third paradigm” for scientific progress, after theory and experimentation.
Grand problems of science and engineering—from modeling the climate a billion years ago to designing safer cars—depend on supercomputing.
How Fast is Fast?
We compare car speeds by measuring MPH (miles per hour). We can compare supercomputer speeds by measuring FLOPS (Floating Point Operations Per Second).
A supercomputer with a speed of one gigaFLOP (one billion FLOPS) produces a new mathematical result every billionth of a second. A teraFLOP is a thousand times faster!
有道翻译
对速度的需求
速度不仅仅是不耐烦的奢侈品。超级计算机通过模拟危险，昂贵或无法在现实生活中尝试的系统来做其他事情无法做到的事情。在理论和实验之后，这为科学进步提供了“第三范式”。
科学和工程的巨大问题 - 从十亿年前的气候模型到设计更安全的汽车 - 依赖于超级计算。
速度有多快？我们通过测量MPH（英里/小时）来比较车速。我们可以通过测量FLOPS（每秒浮点运算）来比较超级计算机的速度。
一台速度为1千兆FLOP（10亿FLOPS）的超级计算机每十亿分之一秒就会产生一个新的数学结果。teraFLOP快一千倍！
Sample Fortran code fragment from a Cray supercomputer
The Cray Fortran language was specially written for supercomputer applications with large computational requirements.
有道翻译
从Cray超级计算机中获取Fortran代码片段示例
Cray Fortran语言专为具有大量计算要求的超级计算机应用程序而编写。
评论