摩尔定律没有告诉我们的事情（之二）

      既然现在我对摩尔定律的持续有效性产生了极大的质疑，那么就让我试着解释一下我们正在为物理学做着何种补偿。

      首先，我曾经说过，在过去的大约十年里，时钟速度所体现的性能并没有按照摩尔定律所预测的那样发展。那我们是如何成功地让其不断加速呢？例如，在有了DSP的情况下，为了增强性能，你是会选择增加时钟还是增加乘法器？

      对此，我们的回答始终是增加乘法器的数目。如果我们回过头看看C1X，每个时钟周期有一个乘法器。当我们发展到C64X的时候，每个时钟周期有四个乘法器。现在，依托新的DaVinci SoC技术，每个周期有八个乘法器。这样，我们就对由物理原因引起的性能下降进行了补偿。然而，摩尔定律并没有预言到这些。

      在电源管理方面，我们有很多可以做。首先，因为温度对电源有反作用，我们必须更好地控制温度。创造出更好的绝缘体是达到此目地的一种方法。另一种办法是节电。有多少次你父母告诉你在离开屋子的时候要注意关灯？这个好习惯可以移植到DSP设计中。我们可以设法让芯片的某些部分在非工作状态时关闭，从而为当时工作中的某些必需功能省电。

      具有讽刺意味的是，通过使用更多的晶体管也可以达到节电的目的。看起来有些不可思议是吧？但是，如果我用一百万个晶体管在1GHz频率下运行，那么在运行5GHz时利用超过两百万个晶体管才能实现同样性能。但是电源损耗就会减少，因为5GHz的晶体管可以在很低的电压下运行。当然，电源损耗与电压的平方成比例。

      所以，我们显然还需要不断创新。在可预见的未来（可能是未来的二十年间），摩尔定律在晶体管数量翻倍（他在1965年所做的预言）的情况下还会发挥作用。但在集成越来越多的晶体管以降低能耗的情况下，我们还需要探索新的出路使性能达到最优化。摩尔之前也没有提醒过我们这个半导体技术的转折。

附：方进英文博客原版

                      What Moore Didn’t Tell Us about ICs (Part Two) [corrected]
                                                           Gene Frantz
                       TI Principal Fellow and Business Development Manager, DSP

Now that I’ve committed the unthinkable heresy of questioning the continued validity of Moore’s Law, let me try to explain how we are compensating for the physics that have caused this stir.

First, I mentioned that performance, in terms of clock speed, has not followed Moore’s prediction for the last decade or so. So how have we managed to keep increasing it? For example, with DSPs, do you increase the clock or do you increase the number of multipliers to increase performance?

Our answer has been to increase the number of multipliers. If we look back at the C1x, we had one multiplier per clock cycle. By the time we developed the C64x, we had four multipliers per clock cycle. And with the new DaVinci SoC, we have eight multipliers per clock cycle. In this way, we compensate for the drop off in performance that is caused by physics. Moore’s prediction never told us about that.

In terms of power, there are a couple of things we are doing. First, because temperature adversely affects power, we have to be better at managing temperature. Creating better insulators is one way of handling this. Another is by being economical with power. How many times did your mother or father tell you to turn off a light when you leave a room? This same good sense can be used in a DSP. We can tell certain parts of the chip to turn off when they are not being used, saving power for the necessary functions at that time.

Ironically, you can save power by using more transistors. Seems counterintuitive, doesn’t it? But if I am using One million transistors operating at 1G, I can get the same performance out of two million transistors operating at .5G. But the power dissipation will be down because the .5G transistors can run at a lower voltage. And, power dissipation is a square function of the voltage.

So, it is clear that we will have to continue to be creative. Moore’s Law, in terms of the number of transistors doubling (his prediction in 1965), will continue to hold true for the foreseeable future (perhaps the next couple of decades). But we will have to continue to find ways to maximize performance while cutting power by using more transistors. Moore didn’t warn us of this twist in the technology either.