13

这是我的一本 CS 教科书中的一个问题。我很茫然。我不明白为什么它必然会导致并行计算。有人想指出我正确的方向吗?

4

16 回答 16

18

Moore's law just says that the number of transistors on a reasonably priced integrated circuit tends to double every 2 years.

Observations about speed or transistor density or die size are all somewhat orthogonal to the original observation.

Here's why I think Moore's law leads inevitably to parallel computing:

If you keep doubling the number of transistors, what are you going to do with them all?

  • More instructions!
  • Wider data types!
  • Floating Point Math!
  • More caches (L1, L2, L3)!
  • Micro Ops!
  • More pipeline stages!
  • Branch prediction!
  • Speculative execution!
  • Data Pre-Fetch!
  • Single Instruction Multiple Data!

Eventually, when you've implemented all the tricks you can think of to use all those extra transistors, you eventually think to yourself: why don't we just do all those cool tricks TWICE on the came chip?

Bada bing. Bada boom. Multicore is inevitable.


Incidentally, I think the current trend of CPUs with multiple identical CPU cores will eventually subside as well, and the real processors of the future will have a single master core, a collection of general purpose cores, and a collection of special purpose coprocessors (like a graphics card, but on-die with the CPU and caches).

The IBM Cell processor (in the PS3) is already somewhat like this. It has one master core and seven "synergistic processing units".

于 2009-02-06T23:12:19.297 回答
13

一个字——

由于无法在当前晶体管水平上散热,工程师们正在利用他们不断增长的晶体管预算来创建更多内核,而不是创建更复杂(和热)的管道和更快的处理器。

摩尔定律根本不是死摩尔定律是关于在给定成本下的晶体管密度。碰巧的是,出于各种原因(如营销),工程师决定使用他们的晶体管预算来增加时钟周期。现在他们决定(因为热量问题)开始使用晶体管进行并行处理,加上 64 位计算和降低功耗。

于 2009-02-06T22:43:31.087 回答
7

摩尔定律描述了由于在电路板上添加更多晶体管而导致芯片性能有效翻倍的趋势。

由于设备的尺寸没有增加(如果相反的话),那么很明显,这些额外的晶体管的空间只是由于芯片技术变得更小和制造变得越来越好而变得可用。

但是,在某些时候,您会到达无法进一步最小化晶体管的阶段。由于产生的热量和所涉及的制造成本,将芯片的尺寸增加到超过某个点也变得不可能。

这些限制需要一种提高性能的方法,而不仅仅是生产更复杂的芯片。

一种这样的方法是在并行架构中使用更便宜、更简单的芯片,另一种是从传统的集成芯片转移到像量子计算这样的东西——根据它的定义,它就是并行处理。

值得注意的是,这个问题的标题更多地与观察到的定律结果(性能提升)有关,而不是实际定律本身,后者主要是对晶体管数量的观察。

于 2009-02-06T22:22:13.267 回答
6

I think it is a reference to the free lunch is over article

basically, the original version of Moore's law, about transistor density, still holds. But one important derived law, about processing speed doubling every xx months, has hit a wall.

So we are facing a future where processor speeds will go up only slightly but we will have more core's and cache to play with.

于 2009-02-06T22:51:58.560 回答
4

这是一个奇怪的问题。摩尔定律不需要任何东西,它只是观察计算能力的进步,它并没有规定它必须以一定的速度增加。

于 2009-02-06T22:21:54.423 回答
2

提高处理器的速度会使工作温度变得非常高,以至于会在您的办公桌上烧出一个洞。芯片制造商正面临着他们无法绕过的某些限制……例如光速。并行计算将使他们能够在不引起火灾的情况下加速计算机。

于 2009-02-06T22:20:44.770 回答
2

晶体管和 cpu 之类的东西越来越小,越来越快。唉,计算的热量和电压成本正在上升。热量和电压问题与实际物理尺寸最小值一样令人担忧。100ghz 芯片会吸收太多电压并变得太热,但 100 个 1ghz 芯片对此问题较少。

于 2009-02-06T22:21:43.910 回答
2

有趣的是,在“必须”并行计算的问题中提出的想法受到了阿姆达尔定律的质疑,该定律基本上说,除非你的程序 100% 是可并行化的,否则拥有并行处理器只会让你走这么远(从来都不是这样)在现实世界)。

例如,如果您有一个程序需要在一个处理器上运行 20 分钟并且可并行化 50%,并且您购买了大量的处理器来加快速度,那么您的最短运行时间仍将超过 10 分钟。这是忽略了所涉及的成本和其他问题。

于 2009-02-21T01:03:06.947 回答
1

The real answer is completely un-technical, not that the hardware explanations aren't fantastic. It's that Moore's Law has become less and less of an observation, and more of an expectation. This expectation of computers growing exponentially has become the driving force of the industry, which necessitates all the parallelism.

于 2009-02-06T22:56:55.367 回答
1

Moore's law says that the number of transistors in an IC relative to cost increases exponentially year on year.

Historically, this was partly due to a decrease in transistor size, and smaller transistors also switched faster. Because you got faster transistors in step with Moore's law, clock speed increased. So there's a confusion that say Moore's law means faster processors rather than just wider.

Heat dissipation caused the speed increase to top out at around 3 GHz for economically produced silicon.

So if you want more cheap computation, it's easier to add more, slower circuits. So the current state-of-the-art commodity processors are multi-core - they are getting wider, but no faster.

Graphene film transistors require less power, and are performing at around 30 GHz, with theoretical limits at around 0.6 THz.

When graphene technology matures to commodity level in a few years, expect there to be another sea change and no-one will care about using parallel cores for performance, and go back to narrow, fast cores. On the other hand, concurrent computing will still matter for the problems it is a natural fit for, so you'll still have to know how to handle more than one execution unit though.

于 2009-02-06T23:08:18.043 回答
0

因为正交计算失败了。我们应该去量子

于 2009-02-06T22:19:42.260 回答
0

Moore's law necessitates parallel computing because Moore's law is on the verge of/is dead. So taking that into consideration, if it is becoming harder and harder to cram transistors onto an IC (due to some of the reasons noted elsewhere) then the remaining options are to add more processors ala Parallel processing or go Quantum.

于 2009-02-06T22:53:32.427 回答
0

摩尔定律仍然成立。晶体管数量仍在增加。问题在于找出对所有这些晶体管有用的东西。我们不能仅仅通过使流水线更深更宽来不断增加指令级并行性,因为证明指令之间独立性所需的电路在你需要证明独立性的指令数量上会显着增加。我们不能仅仅因为热量而不断提高时钟速度。我们可以继续增加缓存大小,但我们在这里遇到了收益递减点。晶体管的唯一用途似乎是将更多内核放在芯片上,这意味着工程师弄清楚如何处理晶体管的工作只是推高了抽象阶梯,现在程序员必须弄清楚该怎么做与所有这些核心。

于 2009-02-06T23:28:26.007 回答
0

我不认为摩尔定律需要并行计算,但它确实需要最终从纯粹的小型化转变。存在多种解决方案。其中一个是并行计算,另一个是协同处理(这是真实的,但与并行计算不同。协同处理是将工作卸载到 GPU、DSP 等专用 CPU 时。)

于 2009-02-21T01:09:44.500 回答
-1

老实说,我真的不知道,但我的猜测是晶体管在某些时候不会变得更小,需要并行分布处理能力。

于 2009-02-06T22:21:38.777 回答
-1

It's because we're all addicted to increasing speed in our processors. Years of conditioning have led us to expect more processing power, year after year. But the physical constraints caused by densely packed transistors have finally put a limit on clock speeds, so increases have to come from a different perspective.

It doesn't have to be this way. The success of the Intel Atom processor shows that processors could just get smaller and cheaper instead. The processor companies will try to keep us on the "bigger, faster" treadmill though, to keep their profits up. And we'll be willing participants, because we'll always find a way to use more power.

于 2009-02-06T23:20:12.230 回答