In the world of computer chips, bigger numbers are often better. More cores, higher GHz, bigger FLOPs, all wanted by engineers and users alike. But there is one measure in semiconductors that is trending right now, and the smaller the better. Enter the Semiconductor Manufacturing and Technology Node (also known as the Process Node).
But what exactly is it and why is it so important? Why is it measured in nanometers, and why are we going all the way down Sesame Street and bringing you this item numbered 10, 7 and 5?
Let's take a trip into the world of process nodes …
But before we dive into things, you will understand things better if you go through our CPU architecture series. In the first part we cover the fundamentals of how processors work and in the second part we examine how engineers plan and design chips.
The key section relevant to this article is the explanation of how computer chips are physically put together. You should read the section on photolithography carefully if you want to understand the manufacturing process in more detail, while in this feature we will focus more on that point that was touched on briefly:
One of the most important marketing terms related to chip manufacturing is feature size.
In the chip industry, the feature size relates to the so-called process node. As we mentioned in How CPUs Are Designed, Part 3, this is a pretty broad term, as different manufacturers use the term to describe different aspects of the chip itself, but not so long ago it was referring to the smallest loophole between two sections of a transistor.
Today it is more of a marketing term and not very useful for comparing production methods. However, the transistor is a critical feature of any processor because groups of them do all of the number processing and data storage within the chip, and a smaller process node from the same manufacturer is highly desirable. The obvious question here is why?
Nothing in the world of processors happens immediately and not even without electrical energy. Larger components take longer to change state, signals take longer, and more energy is required to move current through the processor. Without trying to sound dull, larger components also take up more space, so the chips themselves are larger.
In the picture above we are looking at three old Intel CPUs. Starting from the left we have a 2006 Celeron, a 2004 Pentium M and a really old Pentium from 1995. They have a process node of 65, 90 and 350 nm, respectively. In other words, the critical parts of the 24 year old design are more than 5 Times bigger than that of the 13 year old. Another important difference is that the newer chip contains around 290 million transistors while the original Pentium has just over 3 million; almost a hundred times less.
While the reduction in process nodes is only part of the reason the newer design is physically smaller and has more transistors, it plays an important role in enabling Intel to offer this.
But here's the real highlight: Celeron only generates around 30 W of heat compared to the 12 W of the Pentium. This heat comes from the fact that when electricity is passed through the circuits in the chip, energy is lost through various processes and the majority of it is released as heat. Yes, 30 is a larger number than 12, but don't forget that the chip has almost 100 times more transistors.
So if the advantages of a smaller process node lead to smaller chips, offer more transistors that can switch faster – which allows us to do more calculations per second – and less energy is lost than heat, another question arises: why not? every chip in the world with the smallest possible process node?
Let there be light!
At this point we need to look at a process called photolithography: light is passed through what is known as a photomask, which blocks light in some areas and lets it through in others. There, the light is strongly focused on a small spot as it passes through and then reacts with a special layer that is used in the manufacture of the chip and helps to delimit the different parts.
Think of it like an X-ray of your hand: the bones block the rays and act as a photo mask, while the flesh lets them through, creating an image of the internal structure of the hand.
Image: Peellden, Wikimedia Commons
Light isn't actually used – it's too big even for chips like the old Pentium. You may be wondering how light can be any size on earth, but it is related to wavelength. Light is a so-called electromagnetic wave, a constantly changing mixture of electric and magnetic fields.
Although we use a classic sine wave to visualize the shape, electromagnetic waves have no real shape. Rather, the effect they create when they interact with something follows this pattern. The wavelength of this cyclical pattern is the physical distance between two identical points: imagine ocean waves rolling onto a beach, the wavelength is how far apart the tips of these waves are. Electromagnetic waves have a wide range of possible wavelengths, so we put them together and call it a spectrum.
Small, smaller, smallest
In the picture below we see that what we call light is only a tiny part of that spectrum. There are other familiar names too: radio waves, microwaves, X-rays, and so on.
We can also see some numbers for the wavelengths; Light is about 10-7 feet tall, or about 0.000004 inches!
Scientists and engineers prefer to use a slightly different method of describing such small lengths, and it's nanometers, or "nm" for short. If we look at an expanded portion of the spectrum, we can see that the light actually ranges from 380 nm to 750 nm.
Image: Philip Ronan, Gringer
Go over this article and reread the part about the old Celeron chip – it was made on a 65nm process node. So how can parts be made that are smaller than light? It's simple: the photolithography process didn't use light, it used ultraviolet light (also known as UV).
On the spectral diagram, UV starts at around 380 nm (where the light stops) and shrinks down to around 10 nm. Manufacturers such as Intel, TSMC, and GlobalFoundries use a type of electromagnetic wave called EUV (extreme UV) that is around 190 nm in size. This tiny one Not only does wave mean that the components themselves can be made smaller, but their overall quality can potentially be better. This allows the different parts to be packed closer together, which helps reduce the overall size of the chip.
The different companies offer different names for the scope of the process nodes they use. Intel is quick to refer to one of its newest models to the general public as the P1274, or "10nm", while TSMC simply calls theirs "10FF". Processor designers like AMD create the layout and texture for the smaller process nodes and then rely on TSMC to produce them.
TSMC has been diligently working on smaller nodes (7nm, 5nm and soon to be 3nm) and making chips for its largest customers including Apple, MediaTek, Qualcomm, Nvidia and AMD. At this production scale, some of the smallest structures are only 6 nm in diameter (but most are much larger). To get a sense of how small 6 nm really is, the silicon atoms that make up the bulk of the processor are about 0.5 nm apart, with the atoms themselves being very roughly 0.1 nm in diameter. As a standard figure, TSMC's factories deal with aspects of a transistor that covers less than 10 silicon atoms in width.
The challenge of striving for atoms
Aside from the baffling fact that chipmakers are working to ensure that functions include only a handful of atoms, EUV photolithography has raised a number of serious design and manufacturing problems.
Intel in particular struggled to get its 10nm production to the same level as its 14nm production, and GlobalFoundries had its own problems getting 7nm and smaller production systems up and running. While Intel and GF's problems may not be due to the difficulties of EUV photolithography, they cannot be entirely independent.
The shorter the wavelength of an electromagnetic wave, the more energy it carries, which leads to a greater potential for damage to the chip to be manufactured; Production on a very small scale is also very sensitive to contamination and defects in the materials used. Other problems such as diffraction limits and statistical noise (natural variation where the energy transmitted by the EUV wave is stored in the chip layer) also work against the goal of achieving 100% perfect chips.
Two manufacturing defects in one chip. Image: solid state technology
In addition, there is the problem that in the uncanny world of atoms, current flow and energy transfer can no longer apply according to traditional systems and rules. Holding electricity in the form of moving electrons (one of the three particles that make up atoms) is relatively easy on the usual scale to flow down the conductors.
At the level that Intel and TSMC are working at, this is much harder to achieve because the insulation isn't really thick enough. At the moment, however, the production problems are almost entirely related to the problems inherent in EUV photolithography, so it will be a few years before we can argue in forums that Nvidia handles quantum behavior better than AMD or similar nonsense!
This is because the real problem, the ultimate cause of the manufacturing difficulty, is that Intel, TSMC, and all of their manufacturing colleagues are corporations and aim at Atoms for the sole purpose of generating future revenue. In a research paper by Mentor, the following overview was given of how much more wafers cost for smaller process nodes …
For example, if we assume that the 28nm process node is the same that Intel used to make its Haswell CPUs (like the Core i7-4790K), then the 10nm system will cost almost double per wafer. The number of chips each wafer can produce is highly dependent on the size of each chip, but a smaller process scale means that a wafer can potentially bring more chips for sale, which helps offset the rise in costs. Ultimately, however, as much of these costs as possible will be passed on to the consumer by increasing the retail price of the products, but this must be balanced against industry demand.
The surge in smartphone sales in recent years, along with near exponential growth in smart technology in homes and automobiles, has forced chipmakers to absorb the financial damage of moving to smaller process nodes until the entire system is mature enough to produce high-yield wafers (ie those that contain as few defects as possible) in large quantities. Given that we are talking about billions of dollars here, this is a risky business, and a good part of the reasons GlobalFoundries got out of the race for process nodes.
If this all sounds a little bleak, let's not forget that the immediate future looks bright. Samsung and TSMC have been running their 7nm production lines at healthy margins in terms of volume and revenue for some time, and chip designers are also planning ahead by using multiple nodes in their products.
AMD's chiplet design and strategy, which debuted with their 3rd generation Ryzen CPUs, is being replicated by other chip manufacturers. In this case, AMD's desktop PC processor used two chips made on TSMC's 7nm node and a 14nm chip from GlobalFoundries. The former were the actual processor parts, while the latter processed DDR4 memory and PCI Express devices attached to the CPU.
The graphic above shows Intel's process node changes over the past 50 years. The vertical axis shows the node size by a factor of 10, starting at 10,000 nm. The chip giant has maintained a rough half-life of 4.5 years (the time it takes to reduce the node size by half each time).
Does that mean we'll see a 5nm Intel by 2025? Probably yes, despite their stumbling at 10 nm, they work hard on the way back. Samsung and TSMC have pushed their 5nm production and beyond, so the future looks bright for processors of all kinds.
They are getting smaller and faster, use less energy and offer more performance. They will pave the way to fully autonomous cars, smartwatches with the performance and battery life of current smartphones, and graphics in games that go way beyond anything seen in multi-million dollar films from ten years ago. The future is indeed bright because the future is small.
Note: This feature was originally released in June 2019. We revised and expanded it as it is just as relevant today as it was before. Part of our #ThrowbackThursday initiative.