The art of making chips "smaller"
更新時間: 2021-11-02 14:39:37
In the computer chip world, the bigger the number, the better. More cores, higher GHz, larger FLOPs, for example, are what engineers and users want. But there is one semiconductor measurement that is hot right now, and the smaller the better. That is the semiconductor manufacturing and technology node (aka process node).
But what exactly is it and why is it so important? Why is it measured in nanometers. In this article, we bring you this detail using the numbers 10, 7 and 5.
Let's step into the world of process nodes ......
One of the biggest marketing terms associated with chip manufacturing is feature size.
In the chip industry, feature size is related to what is called a process node. In fact, it's a fairly loose term, as different manufacturers use the phrase to describe different aspects of the chip itself, but not long ago it referred to the minimum gap between two parts of a transistor.
Today, it is more of a marketing term and not very useful for comparing production methods. That said, transistors are a key feature of any processor, as their groups perform all the number crunching and data storage done inside the chip, and very much require smaller process nodes from the same manufacturer. The obvious question to ask here is why?
Nothing in the processor world happens immediately or without the need for power. Bigger components take longer to change their state, signals take longer to transmit, and more energy is needed to transmit power around the processor. Without trying to sound obtuse, larger components also take up more physical space, so the chip itself is larger.
In the chart above, we see three older Intel CPUs. starting from the left, we have a Celeron from 2006, a Pentium M from 2004 and a very old Pentium from 1995. They have process nodes of 65, 90 and 350 nm, respectively. in other words, the critical components of the older design are more than 5 times larger than the relatively younger design. Another important difference is that the newer chips have about 290 million transistors inside, compared to just over 3 million in the original Pentium; hundreds of times less.
While the reduction in process nodes is only part of the reason why recent designs are physically smaller and have more transistors, it does play an important role in Intel's ability to deliver this.
But here's the real problem: Celeron only generates about 30 watts of heat, compared to 12 watts for Pentium. This heat comes from the fact that as current is pushed around the circuitry in the chip, energy is lost through various processes and is mostly released as heat. Yes, 30 is a bigger number than 12, but don't forget that the chip has nearly 100 times as many transistors.
So if the benefit of having a smaller process node is that it will result in a smaller chip with more transistors, allowing for faster switching - which allows us to do more calculations per second - and less energy lost as heat, it does begs another question: why isn't every chip in the world using the smallest possible process node?
To have light!
At this point, we need to look at a process called photolithography: light passes through something called a photomask, which blocks light in some areas and lets light pass through in others. Where it passes through, the light is focused into a small dot, which then reacts with the special layers used to make the chip, helping to determine the location of the individual parts.
Think of it like an X-ray of your hand: the bones block the light and act as a light shield, while the flesh lets it pass through, producing an image of the internal structure of the hand.
Light isn't actually used - even with a chip like the old Pentium, it's too big.
You may be wondering how light on Earth has an arbitrary size, but it has to do with wavelength. Light is something called an electromagnetic wave, a constantly circulating mixture of electric and magnetic fields.
Although we use classical sine waves to visualize shapes, electromagnetic waves don't really have a shape. More importantly, the effects produced when they interact with something follow that pattern. The wavelength of this circular pattern is the physical distance between two identical points: imagine waves rolling onto a beach; the wavelength is how far apart the tops of those waves are. There is a wide range of possible wavelengths for electromagnetic waves, so we put them together in what we call the spectrum.
Small, small, smallest
In the diagram below, we can see that what we call light is only a small part of this spectrum. There are other familiar names for it: radio waves, microwaves, X-rays, etc.
We can also see some numbers for wavelengths; the size of light is about 10 -7 meters or about 0.000004 inches!
Scientists and engineers prefer to use a slightly different method to describe such a small length, which is the nanometer or simply "nm". If we look at the extended part of the spectrum, we can see that light actually ranges from 380 nm to 750 nm.
Go back to this article and re-read the section about the old Celeron chip - it was made on the 65 nm process node. So how do you make parts that are smaller than light? Simple: the photolithography process does not use light, it uses ultraviolet light (aka UV).
In a spectrogram, UV starts at about 380 nm (where light ends) and shrinks all the way down to about 10 nm. manufacturers like Intel, TSMC and Samsung use an electromagnetic wave called EUV (extreme ultraviolet), which measures about 190 nm. This tiny wave not only means that the components themselves can be made smaller, but their overall quality is likely to be better. This allows the individual parts to be packed more tightly together, helping to reduce the overall size of the chip.
Different companies offer different names for the size of the process nodes they use. processor designers such as AMD create layouts and structures for smaller process nodes and then rely on companies such as TSMC to produce them.
TSMC has been working on smaller nodes (7nm, 5nm, and soon 3nm) and making chips for its largest customers, including Apple, MediaTek, Qualcomm, Nvidia, and AMD. at this production scale, some of the smallest features are only 6 nm (though most are much larger than that). To get an idea of how small 6 nm really is, let's take an example. If the silicon atoms that make up the bulk of the processor are spaced about 0.5 nm apart, the atoms themselves are about 0.1 nm in diameter, so, as a ballpark figure, TSMC's fab handles transistors that are less than 10 silicon atoms wide.
The challenge of targeting atoms
Setting aside the incredible fact that chipmakers are struggling to achieve features with only a few atoms, EUV lithography raises a number of serious engineering and manufacturing issues.
The shorter the wavelength of an electromagnetic wave, the more energy it carries, which creates a greater potential for damage to the chip being manufactured; very small-scale manufacturing is also very sensitive to contamination and defects in the materials used. Other issues, such as diffraction limits and statistical noise (the natural variation of energy deposited into the chip layers by the EUV wave transmission), also work against the goal of achieving a 100% perfect chip.
There is also the problem that in this bizarre atomic world, current and energy transfer can no longer be assumed to follow classical systems and rules. Keeping current flowing down closely spaced conductors in the form of moving electrons (one of the three particles that make up an atom) is relatively easy on the scale we're used to - just wrap the conductor in a thick layer of insulation.
At the level Intel and TSMC are working on, this becomes more difficult to achieve because the insulation is not thick enough. For now, though, the production problems are almost entirely related to the problems inherent in EUV lithography, so it will be a few years before we start arguing in the forums that Nvidia handles quantum behavior better than AMD or some other similar nonsense!
That's because the real problem, the ultimate reason behind the production difficulties, is that Intel, TSMC and all their manufacturing confidants are corporations whose sole purpose in targeting atoms is to generate future revenue. A research paper in Mentor provides the following overview about how much wafer cost is needed for smaller process nodes ......
For example, if we assume that a 28nm process node is the same as the one Intel uses to build their Haswell family of CPUs (e.g. Core i7-4790K), then their 10nm system costs almost twice as much per wafer. The number of chips that can be produced per wafer depends heavily on the size of each chip, but going to a smaller process size means that the wafer may yield more chips available for sale, which helps offset the increased cost. Ultimately, however, passing on as much of the cost to consumers as possible by increasing the retail price of the product will have to be balanced against industry demand.
The growth in smartphone sales over the past few years, and the near-exponential growth of smart technology in homes and cars, has meant that chipmakers have had to absorb financial losses from smaller process nodes until the entire system is mature enough to mass-produce high-volume wafers (i.e., those with as few defects as possible). Given that we're talking billions of dollars here, this is a risky business and a big reason for GlobalFoundries to exit the process node race.
If this all sounds a bit pessimistic, we should not forget that the near future does look positive. In terms of volume and revenue, Samsung and TSMC's 5nm lines have been running for some time with high margins, and chip designers are planning ahead to use multiple nodes in their products.
The small chip designs and strategies that AMD debuted on its third-generation Riptide CPUs are being replicated by other chipmakers. In this case, AMD's desktop PC processors use two chips made on TSMC's 7nm node, and one 14nm chip made by GlobalFoundries. The former is the actual processor component, while the latter handles the DDR4 memory and PCI Express devices connected to the CPU.
The chart above shows how Intel's process nodes have changed over the past 50 years. The vertical axis shows node size as a factor of 10, starting at 10,000 nm and going up. The chip giant has followed a rough node half-life of 4.5 years (the time it takes to reduce node size by half each time).
So does this mean we will see Intel's 5 nm process by 2025? Probably yes, although they are stumbling on 10 nm, but they are working their way back. Samsung and TSMC have been pushing production at 5 nm and beyond, so the future looks really good for all kinds of processors.
They will be smaller, faster, use less energy and deliver higher performance. They will lead the way to fully self-driving cars, smart watches with the power and battery life of current smartphones, and game graphics not seen in multi-million dollar movies a decade ago. The future is indeed bright, because the future is small.