Moore’s law has become a sort of mystical talisman in the computing industry. Every few months there is a new round of articles arguing about whether and when Moore’s law will end. In fact, it was only in 2005 when Moore himself finally came out and admitted that his namesake would eventually end. This is puzzling because 1) it was, in fact, arguably already dead in 2005 and, 2) he could have made the same statement in his original paper over 40 years ago because of the fact that it is difficult (never say impossible) to conceive of a transistor smaller than a single atom.
Moore’s Law DIED. Past tense intended. The main hero of Moore’s Law, Dennard’s scaling (= constant power density with shrinking transistors), died a long time ago in 2005. In fact, even this year, in 2013, long past the time when Moore scaling ended, we have articles arguing about if and when Moore’s law will fail. So, for those of you paying attention, I’d like to put this dead horse down for good by showing how Moore’s law is not only dead, but long dead and dead 4 times over.
Will the Real Moore’s Law please stand up?
While the popular notion is to believe, nay demand, it to be one, Moore’s law is not a law. There is no physical, scientific or natural basis for the law. It is an observation: the number of components in an integrated circuit doubles every two years. Part of the problem is that there are several formulations of Moore’s Law and the closely related Dennard’s scaling – an arguably better conjecture of semiconductor chip trends. Moore himself has slowly altered his observation, and in 1975, Moore drastically reformulated the law to state two years over the original one year. And if you have heard 18 months (closer to reality), that’s thanks to Moore’s colleague David House.
But we will not be quibbling about reformulations and the meaning of ‘is’ in the 4 deaths discussed in this series. Each death is a scientific fact based on the most recent, relevant, and popular formulation right before its death.
Death 1: Our Computers Are Getting Dumber
Let’s start with the most popular number of transistors per chip formulation. The green curve in Exhibit 1 below shows Moore’s law, and the green squares show the actual available transistor count. As transistors based on bulk semiconductor properties have shrunk to quantum scales, they start to have size-induced errors called finite size effects. This is a fancy way of saying that things behave differently when the size, edges, and shape of the semiconductor start to matter, but our designs don’t account for it. So size effects become errors. But here is the kicker: an astounding 20 – 30% of transistors in the current generation of processors are never put to use by the user. They must be solely and constantly dedicated to error checking the operations of the rest of the chip.
Exhibit 1: The Green line is the classic representation of Moore’s Law showing the number of transistors on a chip. The Dark Blue line shows the clock speed, or the speed at which the transistors can be used to compute. The Light Blue line shows the power required, and the Purple line is a measure of performance per clock. Source .
What is the point of cramming more components if they are not actually adding to the performance? So, while people are parading about this green curve as proof that Moore’s law is alive and well, the harsh truth is that it died circa 2007 (short red line). We have stopped cramming components that add to performance (an implicit assumption of Moore’s law) onto integrated circuits, and that’s that.
An extremely popular rephrasing of Moore’s law is to track transistor sizes; transistors shrink to half their sizes roughly every two years. In fact, it’s due to this miniaturization that we can pack double the number of transistors than the period before. But transistor sizes long ago went the way of jean sizes. Sorry to break it to you mister, but your waist is not 36 inches. They just noticed that you’re not as likely to buy if they remind you that you’re kinda big. 22nm Technology, 14nm Technology, they’re the same thing. It’s a polite little lie from your friends in the marketing department and do not actually imply 14-nm transistors. Exhibit 2 below shows actual key transistor dimensions. Transistor gate length [yellow] and the metal one-half pitch [orange]—half the distance spanned by the width of a wire and the space to the next one on the dense first metal layer of a chip—have stopped decreasing. The ‘node name’ [red] (a la the marketing lingo) on the other hand seems to keep going down. So Moore’s law is dead in technology but clearly alive and well in marketing. It’s just become a brand, perhaps to keep investors happy.
Exhibit 2: Data Source: Global Foundries. What’s in a name? Key chip dimensions, such as the transistor gate length [yellow] and the metal one half pitch [orange]—half the distance spanned by the width of a wire and the space to the next one on the dense, first metal layer of a chip—have decreased but not strictly tracked the node name [red]. Source .
Death 3: Hidden Hero Slain
The real hero of Moore’s law is Dennard’s scaling. 10 years after Moore, Robert H. Dennard, the inventor of dynamic random access memory (DRAM), showed that as transistors shrunk, the power density remained constant. This meant that a reduction in a transistor’s size by 2 decreased the required power by 4! More importantly, this reduction in required power per transistor allowed them to be run twice as fast without overheating and melting. So thanks to Dennard’s Scaling, a doubling according to Moore’s law not only gave us twice the transistors to compute with, but they could run twice as fast without requiring more power! As you can see in Exhibit 1, finite size effects killed Dennard’s Scaling nearly 10 years ago. That is why when we shop to replace our computers now, we are sadly disappointed that the processor is the same old ~2GHz.
Heating has become such an issue, because Dennard’s Scaling has ended, that there is now vast acreage of useable transistors that must be kept dark so the whole chip does not melt. Let that sink in. A small chunk of the transistors that are being crammed in are not even turned on! Then why exactly is the semiconductor industry going to these lengths to keep Moore’s law alive in the minds of the people?
For the 40 years in which Moore’s law held, roughly every 2 years we received twice the number of processors running at twice the speed for the same price. Now it not only takes longer and longer to double the number of transistors, but every time we manage it, a larger and larger fraction of the transistors are lost to error checking or heating constraints. Additionally, the ones that are left don’t run any faster.
Your new computer is a lot of things. It’s smaller. It’s lighter. It might have a fast solid-state hard drive. But it is not any faster. For the first time in 40 years, it’s prettier not smarter. It’s the end of an era… 10 years ago. Welcome to the present. Let’s get on with the future. You can start by ignoring what technology will get us to 7nm because it doesn’t matter. What matters is which of the many potential usurpers to silicon will provide the runway for another 6 orders of magnitude!
Moore’s law is dead. Long live Moore’s law!
Coauthored with Radhika Dirks for Silicon Valley Story.