One of the greatest advances in the history of technology is the advent of the transistor. The invention of the transistor dramatically changed how we work, how we play, and our imagination of what is possible. From large compute farms that run the internet to personal computers that connect all of us, to cell phones that change the way we communicate, to the incredible potential of artificial intelligence. All of these rely on the constant improvement in power and performance of the transistor.
Hi, my name is Paul Packan, and today I'd like to talk to you about the transistor. I've had the privilege to have worked on 10 generations of Intel technologies, where I work to improve Intel's transistor performance and power to enable the next generation of computers. This ride is nowhere near over. This is Architecture All Access: Transistor Technology. One of the first general-purpose programmable computers was the ENIAC.
It used resistors, capacitors and diodes. But the key was the vacuum tube. It created a way to turn on and off the current flow, which was critical to the computer. This enabled the programmability as well as the amplification of the current, so that the computer could be made. These early computers were very large.
They consumed an extremely large amount of power. And the computing capabilities were, quite frankly, very limited. The vacuum tubes also created a large amount of heat that was very difficult to dissipate. Vacuum tubes had fundamental problems in scaling as can be seen here.
They changed very little over 30 years. This meant that to make a bigger computer, you have to make a bigger number of actual compute parts. This also led to higher power supplies and more power dissipation. It appeared that creating performance computers capable of solving real-world substantial problems would not be possible. Then a breakthrough came in the 1940s with the invention of the transistor.
At AT&T Bell Labs, John Bardeen, Walter Brattain and William Shockley created the first solid-state transistor. Made out of the semiconductor germanium, this transistor enabled switching and application capabilities with the promise of making it much smaller and consuming much less power. This was the moment that changed the world and eventually led to the computer revolution. The transistor is basically a switch.
It's made out of a source and a drain where the current flows into and out of the transistor. The gate is key. The gate controls the flow of current.
The region under the gate is called the channel. Now when the gate is off, current can't flow from the source to the drain. In this case, a negative voltage on the gate will attract positive charges into the channel. This creates a barrier between the electrons in the source and the drain, and no current can flow.
The gate is turned on by applying a voltage to the gate to attract the appropriate charge in the channel. In this case, a positive charge will attract the electrons in the channel. This completes the electrical connection between the source and the drain, and current can flow. The electric fields drive the current flow in the transistors.
They're also responsible for turning the transistor off by creating electrical barriers so no current flows. Although we would like to have no current flow when the transistor is off, some small current still flows. We refer to this unwanted current as leakage current.
The vast majority of today's high performance computers are based on silicon transistors. Silicon is a crystalline material. That means it has a lattice.
It has a periodic structure associated with it. And this is the substrate upon which all transistors are built. The electrons move through the silicon lattice to create currents in the circuit. The more easily the electron can move, the higher the performance. They can also move faster with less power dissipation.
Earlier I showed you a schematic of a transistor and it showed that the source and drain regions there are charges. There's electrons in this case that I showed you. Silicon is a semiconductor, and this is critical to the operation of the transistor. The way bonds are actually constructed within silicon, it has four bonds with atoms. If you put a different atom into that lattice, it needs four bonding electrons. If we add a phosphorus, phosphorus has five bonding electrons.
So, after using the four, there's an extra electron, and that electron is mobile. And that's the electron that we've been talking about. Turns out, though, that you could add an atom that has less than four.
In this case, for example, boron, which has three bonding electrons. There's one electron missing. We call that a hole. It's the absence of an electron. But if you go through it, it operates – it acts – just like a positive charge. So we can create transistors that either have negative charge or positive charge.
We call these nMOS devices and pMOS devices. Both are used to create cMOS, which is a very important technology that are used in computers today. We can use them to create circuits that perform computations. The transistor can be arranged into circuits that have different properties. For example, let's take this first circuit.
If either transistor A or B or both of them are off, no current can flow. However, if we take this circuit: if either A or B is on or both are on, current can now flow. You may recognize these as an AND and OR circuit.
So these two circuits act very differently for the same inputs of A and B. By combining these simple circuits together as shown, for example, in the slightly more complex adder circuit, we can begin to do computations. In this case, adding numbers based on the inputs. Although the simple adder has less than 30 transistors, to do the complex computations of today's CPUs take hundreds of millions of transistors.
Now let's talk about what we need to do to scale the transistor sizes to keep reducing the cost and improving the performance. Once the first transistor was developed, it was quickly demonstrated that the transistor could be scaled to smaller and smaller sizes, using less and less power. By shrinking the size, we could produce more transistors in the same area at the same cost and improve the performance.
This enabled more powerful computers year after year that can solve larger and more complex problems. In 1965, Gordon Moore wrote an article in the magazine Electronics that proposed that the rate of improvement in computer capability would continue to scale over time. That the scaling would result in a doubling of the compute capabilities every 18 months, at the same time reducing the cost per compute.
This cost per compute reduction was key to the economic viability of this new industry. And in fact, this is at the heart of Moore's Law: improving performance while reducing cost. Although not a true scientific law, Moore's Law has been shown accurate for more than 50 years. This law has enabled the performance of computers to dramatically improve over time. Twenty-five years ago, the top computers could enable one teraflop of compute power.
It took almost 10,000 microprocessors in the area of a small house. It also consumed more than 500 kilowatts of power. Just 10 years later, the same one teraflop of compute performance was demonstrated by Intel, that required only 80 CPUs, not 10,000, it could fit on a single computer board – not the size of a small house – and it used less than 100 watts of power. That's about the amount of a light bulb. Okay, you know, I mean the old incandescent ones. When Gordon Moore originally envisioned the scaling of transistors, he recognized that transistors would change their electrical characteristics and would require redesign and reengineering as we made them smaller.
Robert Dennard later derived a methodology for scaling the transistors that became known as Dennard scaling. In Dennard scaling, all dimensions and voltages of the transistor are scaled by the same scale factor. The transistor density, switching speed, and power dissipation per transistor will all improve if this is done.
Let's look at this: if we use a scale factor of 0.7 in the x direction and 0.7 in the y direction, we get a total area scaling of 0.7 x 0.7, which is 0.49, which basically is Moore's Law. Dennard scaling was used as a blueprint for transistor scaling in the early years of transistor development.
In addition to scaling the area to improve the density, we also need to make sure that the performance of the transistors improve. Remember, Moore's Law is really a law saying we're going to give more performance at a lower cost per transistor. Now, if we start with the transistor that has a certain number of dopant atoms in the source and drain region, we need to make sure that we maintain that number in the scaled down device. Otherwise, we're gonna have less charge. If we've got less charge, we’ve got more resistance. The device is going to slow down.
The transistor is not going to be as high performing. So, as we scale this transistor down, we need to keep the same number of dopants. But since the area decreased, the only way to keep the same amount of dopant atoms, the same amount of charge, is to increase the concentration of these dopants. So we have to add more of them.
Now, over the last 50 years, we have added more and more and more dopants to the silicon. And as we keep adding more dopants, they actually begin to get closer and closer together until at some point they really start to interact with each other. Now, we don't want this.
What we wanted was the atom to go into the silicon lattice, either donate an electron or become a hole. But if they get closer and closer together, they don't start interacting purely with the silicon. They start interacting with each other. This interaction will ultimately cause defects in the silicon.
This means they no longer create the extra electron or the hole, so we can't continue to add any more. This means the only way to keep that charge concentration high enough is not to scale the junction depth. But we just said that if we don't scale the junction depth, we're going to increase the leakage. So this is a problem with Denard scaling. Although Denard scaling may be coming to an end, it doesn't mean Moore's Law is. Remember, Moore's Law is an economic statement: more compute power at lower cost.
By reducing the size, we could double the number of transistors every 18 months. This fit into the same area and thereby we doubled the computational capabilities. But this is not the only way to accomplish Moore's Law.
New innovative science was required to continue Moore's Law. You can see how many different elements we're now evaluating and looking at. In the 1980s, we were using a fairly small number.
Hey, note over here, boron and phosphorus that we talked about earlier, are in this chart. They were being used. Now, if we move to the 1990s, what major changes will you see? Well actually, if you look at this, there weren't that many changes. They weren't necessary. We were still scaling using Dennard principles. And so scaling was an engineering feat.
So, now if we look at what atoms are we looking at or evaluating in our technologies in the 2000s, you see how many new atoms and how many new exciting directions are being explored and implemented. We're looking at new materials, new geometries, new device, new fundamental physics. New materials with reduced resistance and increased electron and hole velocities can be used to improve performance also. Now, if you look at this silicon lattice, you can see as I change the orientation, electrons might see a very different path through the actual silicon. The electrons will always go to the lowest energy allowed state.
Depending on where that is in the silicon, which direction they're moving, they're going to have a different speed. They're going to have a different amount of performance associated with them. If we can suddenly move the electrons or holes to a different energy, we can get a different performance out of them. One way to do this is by actually using stress. We can get a feel for this by looking at the lattice and seeing how it changes when we stress the lattice in one direction or more.
These lattice changes, whether we pull out the lattice or contract the lattice, will change what those energies are associated with. And it can either make the electrons or holes move faster or move slower. Clearly, we want to make them move faster. As we talked previously, Dennard scaling is premised on scaling all the different dimensions.
One of the most critical dimensions in traditional transistor scaling has reached some fundamental limits. It's the material that we call the gate dielectric. This material separates the gate electrode from the silicon surface where the current flows. When we apply a bias to the gate electrode to turn on or off the transistor, we don't want any current to flow through this region. This is just going to be leakage current. Now we come to a quantum mechanical problem as we continue to scale the actual dielectric dimension.
We've probably all heard that electrons have both a particle and a wave nature to them. It's easier to understand the particle nature. We're very used to macroscopic items like balls in a pool table colliding with each other. However, at very, very small dimensions, electrons can also act as waves. This was famously shown by Schrödinger's equation. So what does that mean? When you think of an electron as a wave, it has a probability of being in different locations.
When you think of it as a particle, you think it has a very clear location in time and space. So what happens at very, very small dimensions, is that the electron has a probability of being in different locations. Now, if you thin the dielectric to very, very small dimensions, it becomes exceedingly thin on the order of only a few atoms. Now, this tunneling property, this probability of being on the other side of the actual barrier is very high, and we start to get leakage currents. So we've reached these dimensions in traditional transistors.
And this solution to Schrödinger's equation shows us that our drive currents, our currents that are associated with the leakage, are going to go up exponentially the further we scale the thickness of the dielectrics. Another problem for Dennard scaling. But let's see if we can find a paradigm shift to keep Moore's Law going. If we look at the reason that we needed to scale the dielectric thickness in the Dennard scaling, it was to maintain the charge in the channel. We talked before about how the charge is very important.
This is the current that we're going to see. This is the performance. The equation for charge is: charge is equal to capacitance times voltage. This shows that the change is proportional to the capacitance. So what is the capacitance? Well simply put, it's the dielectric constant divided by the thickness of the material. So now we see why Dennard scaling needed to scale all dimensions.
It needed to scale the thickness, to improve the capacitance, to maintain the charge. So the thinner the dielectric, the higher the capacitance. But another way to increase this capacitance is by increasing the dielectric constant. Now, to do this, we need to find a new dielectric material.
If we could, we can make the dielectric material thicker to remove all this leakage issue associated with Schrödinger's equation and quantum tunneling and still increase the charge and maintain Moore's Law. It's not simple to find a new dielectric material that has all the process requirements and the properties required for high performance transistors. It's not just finding a high-k dielectric. It's finding one that works with silicon, that works with the electrons and holes, that actually enables high performance. We were able to find a new system of materials that has been adopted by the industry that is able to deliver this high performance and low leakage.
And this continues Moore's Law. Dennard scaling may have ended, but you can shift to a new paradigm. And this new paradigm can continue to enable Moore's Law – just in a different way. By creating a new transistor with different geometries and characteristics, we can overcome the issues of Dennard scaling.
This is a planar device. We have a gate, as I've shown you on top. We've got a source and a drain on each side. As we actually bias the gate, we can turn off the current near the surface. But deep into the bulk, the leakage currents can still occur. So now we've created a different type of device.
It's called a FinFET. This new transistor was introduced about 10 years ago by Intel. In this transistor, it fundamentally has changed the shapes. The FinFET transistor improved the leakage by confining the transistor channel between two plates, the top and the bottom.
This makes it much easier to control the leakage current, as now you don't have just the front side, but you have the front side and the back side. This device does not require scaling of the source drain regions because I have control with two channels. So, although Dennard scaling may be over, Moore's Law certainly isn't. Many new innovations, including transistor strain, high-k dielectrics, FinFET devices have enabled Moore's Law to continue. There are many new ideas and innovations that will continue well into the future. In the end, Moore's Law is continuing.
It's changed. It's morphed. And as Gordon Moore himself told me, many people have predicted the end of Moore's Law, but it's still going strong. Thanks for joining me. My name is Paul Packan. This has been Architecture All Access: Transistor Technology and Moore's Law.
2021-10-22