You can't read more than a couple of articles about upcoming CPU's or chipsets without running into one claiming that Moore's Law is coming to an end. Moore's Law is usually (and incorrectly) stated as "microprocessors double in speed every 18 months". They then claim that with the end of Moore's Law, we'll stop seeing improvements in computer technology.
The first thing wrong with Moore's Law is that it's not a law, but an observation. Laws are unbreakable, but observations are. Second, Gordon Moore said that it was the transistor density that was doubling, not the speed. And third, he noticed that it was doubling every two years, not eighteen months. And fourth, the observation was made regarding silicon semiconductor technology.
Interestingly, Moore made the observation in 1965 using data from 1958 to the then-present. From there, the Law has actually held up pretty well for the last fifty years. The reason for this may simply be a case of the prediction leading the technology. If transistor densities have doubled like clockwork and you know your competition is also working on the next doubling, then you are well-motivated to make it happen.
Even so, Moore's Law is coming to an end. Silicon transistors can only get so small; single atom transistors are as absolutely small as you could ever hope to get and for silicon, that's about 0.25nm. Right now, Intel's latest chips use 22nm technology (which is half of the feature width). So, there is a limit and we are approaching it. There is no way that transistor densities can keep doubling for the next 20 years without having to have transistors smaller than a single atom.
But I'm not worried and here's why. First, Moore's Law makes a prediction about transistor size, not processing speed. Current chips have over a billion transistors on them and when you have that much wiring, there is invariably room for improvement. In fact, the current battle between Intel's x86 instruction set and ARM's RISC instruction set is evidence of that. If there was a "right" way to hook up a billion transistors, Intel would do it and be done with it. But the fact that they have to compete with ARM and AMD shows that there is still competition in the architecture regime.
Also, even for a fixed transistor density, you can increase the number of transistors by increasing the chip area or going three-dimensional. All chips today are made in nearly flat 2D layers, but research has been done into layering transistors. If this were accomplished, you would have an overnight doubling of transistors on a chip. And since each layer is only a few nanometers thick, you can have thousands of layers before the chip becomes appreciably thicker. Of course, there are a lot of challenges in this arena, specifically in heat dissipation and signaling, but it is a way forward.
Moore's Law is also an observation on silicon semiconductor integrated circuit technology. The original electronic transistors were vacuum tubes, which were then replaced by individual transistors. Those were replaced by silicon integrated circuits which we use today. It is inevitable that someday we will move away from silicon to something better. We may figure out how to make transistors at the subatomic level, or quantum computing might finally mature. But most likely, it will be some technology that we haven't even heard of yet that will replace it.
In the short term, Moore's Law looks like it still has a bit of life left in it. I expect that within this decade we will see the first signs of a transition away from silicon to a similar technology (such as gallium-arsenide transistors). Then, shortly thereafter we'll see a transition away from the integrated circuit technology that has been the center of computing and technology for more than half a century.
If you feel like checking it out for yourself, Intel has made a copy of the original 1965 paper available on it's website.
No comments:
Post a Comment