How Dead is Moore's Law? (Patreon)

[This is a transcript with links to references.]

My family’s first personal computer was a Commodore 128. It had 128 kilobytes of RAM and a display resolution of 320 times 200 pixel. The laptop I currently use is a few years old but has 16 gigabytes of ram and a screen resolution that outperforms my rapidly aging eyes. There’s no doubt that computing power has seen the most remarkable progress ever since the production of the first microchips.

And yet in the past 10 years or so, tech specialists have repeatedly voiced concerns that this progress will soon hit the wall. Miniaturisation has physical limits, and then what? Have we reached these limits? Is Moore’s law dead? That’s what we’ll talk about today.

Up until now, the most important factor driving the increase of computing power has been the miniaturisation of transistors. A transistor is basically an electronic switch. It’s the building block of logic gates, the things that computers compute with. Transistors are to computers what baristas are to Starbucks. They get things done.

Unlike baristas though, the size of transistors has shrunk dramatically since they entered the consumer market. The first transistors in the 1950s had a size of a few centimetres. Today they’re a few nanometres. That’s about ten million times smaller. If you shrunk a barista that much, they’d end up being about the size of a virus.

In 1965, Gordon Moore, one of the brains behind Intel, noticed something interesting: The number of transistors on a microchip was doubling roughly every two years. This became known as Moore’s Law. And as the size of transistors shrank, the number of transistors that could be put on a single chip increased, from roughly 5000 in the 1970s to more than 50 billion today.

It’s not a law of nature though, and sadly also not a law that you can sue over if someone breaks it. Indeed, Moore’s law was always doomed to fail at some point. You see, an exponential increase can’t continue indefinitely. If you keep shrinking transistors, eventually you’ll reach the size of an atom. And an atom is an atom. It’s not a transistor.

Even before that, other physical limits come into play. As we cram more and more transistors onto a chip, it becomes more difficult to get rid of heat. And quantum effects begin to play a role, too, most importantly quantum tunnelling. The issue is, transistors work by switching electric currents on and off. But if you make transistors too small, then electrons can tunnel into places where they shouldn’t be, currents leak, and performance drops.

Have we already reached the end of Moore’s law? Depends on who you ask. Intel thinks we haven’t yet reached it. Nvidia thinks we have. The reason it’s so complicated is that Moore’s law isn’t just about transistor size, it’s about the number of transistors on a chip. That number also depends on how you arrange and connect the transistors. And that in return depends on how you produce them. And that in return depends on how much you’re willing to invest. So in the end, as usual, it’s all about money.

The trouble is, the production of today’s most advanced logic devices require a whopping 600 to 1,000 steps, a level of complexity that’ll soon rival that of getting your travel reimbursement past university admin. Why has it become so difficult to produce a microchip? It’s because as transistors reach the nanometre scale, they become increasingly prone to defects, and that requires more elaborate methods for fabrication and monitoring. Many of those production steps are checking and re-checking for physical defects. It’s called “critical line monitoring”.

There are several ways producers do this. One is e-beam lithography. For this, one sends out electrons to the microchip. They scatter and bounce back to a detector, which can locate physical defects. Another method, called optical inspection shines a large spectrum of light onto the microchip, and again image is investigated for defects. Depending on the number of defects and the stage of production, the chips will either be discarded, or a repair will be attempted. Either way, it costs more money.

Making high-tech semiconductors has never been cheap, but those extra steps could make further miniaturization prohibitively expensive. In the past, each new chip tech wave has required roughly 10 to 20 billion dollars per factory. For major manufacturers like TSMC and Global Foundries, investing that much money has been tough. For smaller companies, it’s been impossible.

This is why fewer and fewer companies are left to push forward with the miniaturization game. Back in 2001, 19 companies were manufacturing chips with cutting-edge transistors. Fast forward to today, and we’ve only got four left standing: GlobalFoundries, Intel, Samsung, and TSMC.

Those other 15 companies? They’ve made a strategic shift to churn out older, so-called “legacy” chips, that are still widely used but can be produced more cost-effectively. And honestly, it’s a pretty savvy move. Legacy chips are everywhere – they’re in cars, home appliances, medical gear, automation systems, you name it. And there’s a long line of companies ready to snap them up.

So we’re down to four companies pushing the limits. In 2015, the Semiconductor Industry Association predicted that it wouldn’t make financial sense for companies to keep on shrinking transistors in microprocessors the traditional way after 2021. But here we are in 2023, so what is happening?

There are a couple of new technologies on the horizon that are pretty much ready for application. We’ll look at those first, and then we’ll talk about what’s going on in laboratories.

One of the current strategies to push the limits of Moore’s law is specialization. Rather than relying on *one type of processor, you use several ones, each for specific tasks. This allows computers to run faster and more efficiently.

It’s called “heterogeneous computing” and has already been going on for quite some time. For example, your computer almost certainly has two different types of processing units. There’s the central processing unit, CPU for short, and the graphics processing units, GPU for short. GPUs have been around since the 1990s and were pioneered by Nvidia. They’re extremely good at parallel processing and are now being used for many other things besides graphics.

This idea can be pushed much further, and indeed there are lots of other processors in use already, such as NPUs, neural processing units, that are used for machine learning in particular. This allows a more dedicated task-assignment which makes the system faster and more efficient. Basically, you get more compute out of the same space, even if that hurts your English processing unit.

The nice thing about heterogeneous computing is that it doesn’t require a large break with the existing production technology, so it’s a shift that can be done at low cost. This is good news both for companies and for consumers. Some experts have estimated that heterogeneous computing might allow companies to replace what’s currently a high-cost device with one that’s one tenth of the cost, with better energy usage. We’re already seeing this approach in devices like the Apple M1 chip that combines CPUs with GPUs and several other things, all on one chip!

Another fairly obvious improvement is making the process better suited to mass-produce nanoscale structures at precision.

One recent advancement is Extreme Ultraviolet Lithography. This technique uses ultraviolet light with a wavelength of about 13 point 5 nanometres to etch patterns onto silicon wafers. This is significantly shorter than the wavelength used in traditional lithography techniques, and allows for the creation of much smaller features.

The idea had been around for decades, but it’s really only been in the past few years that it’s moved into mass production. Major chip manufacturers like TSMC, Samsung, and Intel are now integrating ultraviolet lithography into the manufacturing processes of their smallest chips.

However, it’s a ‘complex process that still poses many challenges. For example, the extreme ultraviolet light is absorbed by almost everything, including air, so the entire process must take place in a vacuum. At the moment, it’s also slower than the older techniques. Plus, the equipment required for ultraviolet lithography is incredibly expensive. The machines currently cost 200 million dollars a pop, and the next generation will cost as much as 350 million.

So these changes are already on the way, but what’s next? The currently most promising new technology ready for application is to fully move into the third dimension. Before we can see what that means, we need to talk a little more about how a transistor works.

As I said, a transistor is basically a switch. Just like your everyday switches that you use to turn lights on and off, transistors control the flow of electricity. A transistor has a “source”, where the electricity comes from, a “drain” is where it goes, and a “gate” that’s like a door that controls whether the electricity can pass from the source to the drain. The part where the electricity flows from the source to the drain is called the channel.

Transistors used to be printed on flat surfaces, but 3-d transistor designs have been around for about 20 years. In those transistors, called finFETs, the conducting channel isn’t flat on a surface, but it stands out. This improves cooling, reduces power consumption, and speeds up operations.

These transistors have been widely used now for more than a decade. You can also put several of those channels next to each other. A new transistor technology which is already under development is called RibbonFET. For this one turns the channels by 90 degrees so that the gate wraps around them. But the big upcoming change is to stack transistors on top of each other, so that you pack them in volume, rather than on area.

Intel, for example, has been developing a 3D stacking technology, which allows putting  different types of chips, or “chiplets,” in a single package on top of each other. Intel is on its fourth iteration of this technology, called Foveros Direct, and they say it’s increased transistor density by 6 times over the previous iteration. Intel hopes to bring this technology to mass production this year. TSMC has its own version of 3D stacking known as Chip-on-Wafer.

But, you saw it coming, the idea also has its problems. Transistors get hot as they run, and packing them closely makes it difficult to get rid of that heat. It’s like when you pack too many people into a small room - it starts to get really warm. Now imagine if those people were running a marathon in that room - it would get hot, fast, and not in a good way.

No problem you might say, just cool them. Unfortunately, the currently used cooling methods, just aren’t up to the task.

This is why researchers are looking into more advanced cooling methods, like actually running tiny channels of coolant right through the chip stack itself. But that comes with its own problems, like the risk of leaks and making the manufacturing process even more complicated. Besides, the cooling equipment might make the system larger, which would defeat the purpose of making the chips smaller in the first place.

Despite the challenges, the commercial interest in 3D stacking is at an all-time high and governments are supporting the technology too. For example, just at the beginning of the year, the US-agency DARPA created a grant program called the Minitherms3D, that will support research on thermal management of 3D stacks. Though Minitherms sounds to me like the name of a home sauna company.

That’s the changes which are already on the way, now let’s have a look at what new things might come out of the laboratory.

Transistors have traditionally been made by adding impurities to silicon. But some new materials might be better suited to shrinking them down.

The first one is graphene that could be used for the conducting channel on the transistor. Graphene is a single layer of carbon atoms, set up in a honeycomb pattern. It has some qualities that make it seem ideal for microchip applications, most importantly it gets rid of heat very efficiently. So it’s rather unsurprising that the idea of graphene transistors has been around ever since the material was discovered in 2004.

Unfortunately, graphene doesn’t have a good band gap, which means it can’t fulfil the function of a transistor to act as a switch, at least not the way it is. This can be remedied by several tricks, such as adding a second layer onto the graphene, or applying strain to it. Using these methods, some graphene transistors have been built and been shown to work. However, they are difficult to produce on a commercial scale which makes it expensive. So, these things are unlikely to end up in your phone any time soon.

However, you can roll graphene up to tiny tubes, called carbon nanotubes, and these get some new properties. Depending on how you twist the tubes, they’re either conducting or semiconducting. You can make a tiny transistor out of those, by using a number of parallel nanotubes, each just one nanometer in diameter, though several other configurations are also possible.

In the lab, carbon nanotube transistors are currently all over the place, and some manufacturers, like Intel and IBM, have also taken an interest. But again, the problem is that the stuff is difficult to produce at commercial scale and therefore expensive. In 2020, researchers at MIT repurposed a commercial silicon manufacturing plant to create carbon nanotube transistors. But so far these techniques haven’t made their way into large-scale production.

There are a lot of other materials that researchers have tried, which have one or the other advantage over silicon, such as gallium nitride, germanium-tin alloys, boron arsenide, molybdenum disulfide, and a couple of other things I don’t know how to pronounce. Let me just say that while those have been tested in laboratories and some look promising, none of them are anywhere close to getting on the market.

Let’s then look at some completely new computing paradigms. On the most basic level, all a digital computer does is shovel around bits encoded in two states, usually called 0 and 1. Transistors do the shovelling. So far we’ve talked about improving the shovel, but you can also change the things it’s shovelling, the bits.

At the moment, the bits are processed by moving around electric currents. But there are other ways to do it, and some of them might lend themselves better to miniaturisation. For example, any quantum-bit that’s used in a quantum computer can also act as a bit, we just need to find a way to miniaturize how we process them.

The most promising approach to this might be quantum dots, that are nanoscale sized structures in semiconductors. Quantum dots basically embrace the quantumness rather than trying to avoid it. Transistors for quantum dots have been built but for the time being they’re not anywhere near technological applications.

It's a similar story for photonic computing. Photonic computing, as the name suggests, uses light to compute. The advantage of this is that the quanta of light, the photons, are small and very, very fast. The disadvantage is that light doesn’t interact with itself. That’s why you can see me, so I don’t want to complain too much about it, but it’s unfortunate if you want to calculate with it. While transistors for single photons have been built using microscopic wires, this is still very much on the laboratory stage. It’s an incredibly active research area though, and I talk about this a lot in my weekly science news.

Another new hardware option that researchers are pursuing is called spintronics. It uses the spin of electrons in a material as the basic units of calculation. The advantage is that flipping the spin of an electron uses much less power and is much quicker than having to moving around the electrons, as in present-day computers.

But there are still many obstacles to overcome to bring spintronic computing on the market. The biggest hurdle is temperature. Most experimental spintronic transistors need to be cooled to temperatures near absolute zero, and that’s more effort than most consumers are willing to put into a text message.

There are a few exceptions though. In 2022, a group of American researchers studied a particular type of perovskite crystal that they claimed would be a good candidate for a spintronics transistor working at room temperature. Another group of researchers achieved a similar thing by combining graphene with a material called a van der Waals heterostructure. They were able to show that they could switch electron spins at room temperature without magnetic fields. So that’s a promising start but there’s a long way to go.

In summary, rumours of the death of Moore’s law have been greatly exaggerated. It might not continue in its original form, but its spirit of progress is as relevant as ever. Many of these experimental research methods are still impractical, cost-prohibitive, or face scaling difficulties. There’s much work to be done for sure before you’ll have them in your phone. But scientists and engineers are still pushing boundaries, breaking new ground, and making cool stuff happen in computing.