主题：【知识】现代的CPU是如何制造出来的 -- Highway

都是英文的，大家就凑合的看吧。

While it's nice to just look at the finished product itself, sometimes it's useful to go back and look at how it's made. Especially today in the silicon industry, where both major players in the x86 desktop market are having issues with their top end products. Another large member of the industry, IBM, is also finding the going at 90nanometer a lot harder than they predicted. Today on Sudhian, we'll take a look at just how a processor goes from essentially sand to a fully functioning integrated circuit, and all the steps in between.

The Basic Materials

Everyone is generally aware that the major component of CPUs today is silicon. In basic terms, that's what sand on the beach is too. Of course, there are a bunch of other elements in there as well, which is why you don't see Intel or AMD just buying up beachfront property for use in their fabs. Instead, very carefully selected silicon, only the purest stuff available is used. You don't make some of the most complicated manufactured products on the planet with crummy, cheap materials. Especially considering what it is they do to that silicon before it even comes close to being the final product that's currently heating your case.

Another basic material used in the process of creating a processor is metal. You'll see later on where this comes in. While up until recently, aluminum has been the metal of choice for inside the processor itself, copper is taking over for modern processes. There are a few reasons for this. Aluminum is more prone to electromigration in current high power designs compared to copper, where the individual atoms move out of place creating holes in the connections. As you can guess, holes don't conduct electricity very well. This is why many "Northwood" Pentium4's suffered from "Sudden Northwood Death Syndrome" or SNDS when overclockers first started applying massive amounts of voltage to them. That was Intel's first experience with using copper interconnects, and it obviously needed some tweaking. Copper interconnects can also be made much smaller, an important fact when we start talking about feature size in nanometers. Lastly, copper has less resistance, allowing electrons to pass through quicker.

There are also many different chemicals used for creating the designs in the silicon itself, and doping it to create different properties. This we'll explain as we get to it.

Preparation

After amassing the raw materials required, some of them require a large amount of preparation. One of those is the main component, silicon. First, it has to be chemically purified, and turned into electronic grade material. For this to turn into the base of your integrated circuit, you also need it to be a uniform solid, instead of essentially grains of sand.

This is done by melting down the silicon, and putting it into a large, heated, quartz bucket. From this, very carefully a first seed crystal is inserted into the melt by a wire. From high school chemistry, you should know that many solids follow a crystalline structure. Silicon is one of these. In order for a high performance microprocessor to come from this silicon, the whole base must be defect free, and be a single crystal. By being rotated and pulled out very slowly, an ingot of silicon is created, the whole mass following the orientation of that first seed.

Up until recently, these ingots were 200mm in diameter. Now however, Intel and others have invested in creating 300mm wide ingots. Making a bigger ingot while retaining those necessary properties obviously is more difficult, but has been solved by having enough money thrown at it. Building a current 300mm fab is approximately 3.5 billion dollars for the output in wafers that is required by Intel, along with the equipment to make rather complicated microprocessors (compared to simpler DRAM). A similar 200mm fab costs around 1.5 billion.

The difference in costs is due to being on the bleeding edge of technology. Since a change in wafer size only occurs once a decade (or more), being the first to try it out does get expensive. The payoffs you'll learn later show why this is money well spent. There are other methods for making silicon ingots, but the CZ method described above was the one taught in my electrical engineering courses.

After the ingot has been created, and ground into a perfect cylinder, it needs to be sliced up into the wafers upon which the process of making a CPU is applied. The thinner the wafer, the more processors you can make. After they've been cut, they are then polished to a mirror finish, and checked for warping and other problems. Again, quality checking here is paramount to achieving a working processor at the other end.

The new wafers are then taken and doped appropriately for the type of transistors that will be made out of them. Doping amounts to depositing other elements into the space between silicon atoms. This is what causes silicon to be the "semiconductor" that it is. Transistors today are made from "CMOS" technology, or Complementary Metal Oxide Semiconductors. Complementary means the interaction of "n" and "p" MOS circuits. N and P as it does in most electronics, defines negative and positive. In most cases, the wafer is doped to be a "p" type substrate, which is necessary for nMOS circuitry, as this type of transistor is the more efficient space and energy wise. In most designs, you try to limit the number of pMOS transistors used. These are grown in later during fabrication with specific "n-wells" put into the oppositely doped p substrate where they are needed.

So now we've got a wafer, and it's been doped appropriately. Now each wafer gets put into a very hot furnace, for a controlled length of time to develop a silicon dioxide layer on it's surface. By closely monitoring the temperature, atmosphere and length of time in the furnace, a specific depth of SiO2 is created. On Intel's 90nm process, this gate oxide is amazingly only 5 atoms thick. This layer is part of the "gate", the part of the transistor that controls the flow of electrons from the "source" to the "drain". Depending on the voltage found at the gate, electron flow can be constricted, regardless of what is demanded by the difference in source and drain voltage. These three parts are aptly named considering their functions.

The last part of the preparation is covering the SiO2 in what's called "photo-resist". This is applied in another controlled, uniform layer. This chemical is sensitive to light once it dries, and can be chemically dissolved after exposure separate from the part that is not touched by light.

Masking

This is one of the more complicated steps in CPU manufacture currently. Why? Photolithography is the process of using light to expose certain parts of the photo-resist layer, and change it's chemical properties. This technique is pushing the limits of what can be done with light. Even using short wavelength UV light and large lenses, the "masks" or "stencils" still manage to blur features together. Remember the scale we are talking about here. Each mask itself is an incredibly complicated device. Every one requires on the order of 10GB worth of data to properly describe it, and there's 20 or more of them (one for each layer) for each processor design. For each mask, think of a map of New York City and it's surrounding area. Then drop it's size down to one hundred millimeters square. And while you are at it, don't forget to connect each map to the others.

Once these masks have been created, they are put over the wafer in turn. The short wavelength light is passed through by the holes in the quartz mask, to the photo-resist material covering the wafer. Then both the light and the mask are removed. A chemical etch is used to remove the exposed photo-resist, and with it the silicon dioxide immediately below it. This gives access to the layer of silicon below.

Doping

After the remaining photo-resist is removed, you are left with ridges of silicon dioxide and exposed silicon underneath all across the wafer. From here, another layer of SiO2 is created, and a layer of polysilicon is added, with another layer of photo-resist. Polysilicon is the other part of the "gate" mentioned earlier. While this used to be metal (hence the name "metal oxide semiconductor"), polysilicon allows for the gate to be created before the source and drain are which aides in alignment of the transistor. Again, the photo-resist is exposed to light through a mask, which determines eventually where the polysilicon will be left over. Another process of etching, and we have our gate, along with exposed silicon. From here, the exposed silicon is bombarded with ions to create an "n" or "p" well. This doping creates diffusions which change how the silicon passes electrons from one end of the eventual transistor to the other. Each individual transistor has a source and drain, with a gate in between. The source and drain are each one of those now heavily doped pieces of exposed silicon.

Wash, Rinse, Repeat

From there, you go on and repeat layering, adding a layer of silicon dioxide, then etching it with photo-resist. This creates a 3D structure that is your eventual processor. In between every few layers, metal is deposited to carry the electrical connections. Today, Intel's P4 processors use 7 metal layers, while AMD's Athlon64 uses 9. The number of layers of metal are determined by the layout, and how much current must be carried.

Testing, testing, and more testing

Following the weeks required for the initial wafer to be turned into layers and layers of metal, silicon and other materials, it's time to see what developed. The wafer itself is tested for electrical properties, to see if any errors were made, and if possible what step they occurred at. Following that step, each die on the wafer is tested individually to see if they meet initial specifications.

Wafers that pass as a whole are then cut up into the individual processors. From the initial testing, ones that are known to be dead are tossed aside. Then each die is attached to it's packaging, which allows it to seat and interact with the motherboard. As well, most recent processors from both Intel and AMD are then covered with an integrated heat spreader. Each processor is tested again at this stage for complete functionality. This is where some "binning" occurs. Some wafers work out to produce processors that run at a higher speed than others, and are binned initially there. Even from less "perfect" wafers though, you do get some that are capable of running at a higher speed grade. Considering the disparity in yields and price, binning can make more efficient use of turning processors into revenue. Finally, certain parts of a processor may not function properly. If this area is in the cache (which takes half or more of a CPUs die area), some times it can be routed around and masked off. This means you can still sell the CPU for a profit, just as a Celeron or Sempron with less enabled cache onboard.

Once the CPUs have been boxed up, many are still tested again to ensure all the previous work has not been in error, and they have been binned properly to meet specifications.

We hope this answers some of your questions on what goes into making a processor. After having done this basic look into the method of building a processor from the ground up, we can go on into what's been done in recent years to increase the speed of transistors themselves, as well as reduce power lost as heat.