What Moore’s Law Says
Gordon Moore made his original observation in 1965:
This can be interpreted in a few ways, but it implies two things. First, (at the time) the most basic Integrated Circuit (IC) would double in transistor density every year. Second, that this would also be true at the lowest cost level. So if the cost to manufacture an IC of a given size remains stable over time (taking inflation into account), this would effectively mean that the cost per transistor would halve every two years.
This is a startling level of exponential growth demonstrated by the “wheat and chessboard problem” where if you put one grain of wheat (or rice) on the first square and then double the amount for each successive square, you’d be up to well over 18 quintillion grains by square 64!
Moore later revised his observation to extend the time to once every eighteen months, and then eventually once every two years. So while transistor density is still doubling, the pace seems to be slowing down.
It’s Not Actually a Law
Although it’s been nicknamed Moore’s “Law,” it isn’t a law in the proper sense of the word. In other words, it’s not like a natural law that describes how things like gravity work. It’s an observation and a projection of historical trends into the future.
On average, Moore’s Law has held up since 1965, and in some ways, it’s a benchmark for the semiconductor industry to tell roughly whether it’s on track, but there’s no reason why it has to be true, or remain true indefinitely.
There’s More to Performance Than Transistor Density
The transistor is the fundamental component of a semiconductor device, such as a CPU. It’s from transistors that devices such as logic gates are built, allowing for the structured processing of data in binary code.
In theory, if you double the number of transistors that you can fit into a given amount of space, you double the amount of processing that can happen. However, not just how many transistors you have but what you do with them counts. Microprocessors have received many advancements in efficiency, with specialized designs to accelerate specific types of processing, such as decoding video or doing the specialized math needed for machine learning.
Shrinking transistors generally also means reaching higher operating frequencies while using less power for the same amount of processing power from a previous generation. Moore’s Law is confined to transistor density, but the relationship between transistor density and performance isn’t linear.
What Do You Mean “It’s Dead”?
Over the years, the phrase “Moore’s Law is dead” has been uttered several times, and whether that’s true depends on your perspective. Transistor densities are still doubling, but at a slower pace as Moore has revised the time frame several times now.
The reason why some contend that the law is dead isn’t that transistor density isn’t still doubling, but that the cost of transistors isn’t halving. In other words, you can’t get twice the number of transistors for the same money after a doubling cycle anymore.
One important part of why this is happening is because we’re approaching the limits of how small we can make transistors. At the time of writing, 5nm and 3nm manufacturing processes are the current and next generation of technology. As we push towards the ultimate limit of what’s possible, the number of problems and the cost of overcoming them are both likely to increase.
However, just because transistors may not be halving in price the way they used to does not mean that performance isn’t doubling or halving in price. Remember, transistor count is only one part of performance. We’re achieving higher clock speeds, fitting more cores into a single processor unit, doing more with our transistors, and creating novel silicon that can accelerate specific jobs such as machine learning. In this expanded sense, Moore’s Law still has life in it, but in its original form, it’s on life support.
Moore’s Law Has to Die Sometime
No one ever believed that Moore’s observation about transistor density and cost would hold true forever. After all, the exponential plot would eventually trend towards infinite transistor density and computing performance. As far as anyone knows, that’s not actually possible, and it’s particularly unlikely to be possible using semiconductor electronics as we know them today.
There are already numerous challenges with the tiny components in modern processors struggling with unwanted quantum effects. At some point, you can’t keep electrons inside your tiny circuits anymore, so trying to make things smaller hits a brick wall.
At that point, it may be time to move to another type of computing substrate, such as photonics, but there are likely myriad ways to get more performance from semiconductors that don’t involve making transistors smaller.
We’re already seeing cost-effective ways to build large processors from multiple smaller processors, such as AMD’s chiplet designs or Apple’s strategy of gluing their baseline chips together to make mega CPUs that operate as if they were one system. There’s potential in the idea of building CPUs with 3D circuits, with layers of microchip components that communicate vertically and horizontally.
While the ultimate limit of transistor density seems to get closer and closer every day, the true limit of achievable computing power is still an open question.
RELATED: Huge Supercomputers Still Exist. Here’s What They’re Being Used for Today
