Designing and manufacturing a modern CPU is a huge project. It requires both backward compatibility and an understanding of where PC workloads are going in the future—a delicate balancing act made more difficult by the huge engineering staffs and massive dollar outlays involved. Let’s take a look at the steps needed to build a Core i7 or AMD Phenom II processor.
Before the manufacturing plant starts churning out chips, there are a few critical preliminary steps. Prior to the first circuit being laid out or the first simulation run, the designers need to know exactly what it is they’re designing. This phase takes input from many sources. Marketing gets involved, with predictions of what users will need when the CPU actually ships, usually two to four years in the future. Engineering and performance teams feed in billions of traces of actual applications being run on current-gen CPUs, so the designers can see how existing CPUs perform under real-world conditions.
After the specification phase, the design phase begins in earnest. Design involves creating a design document, validating the design with simulations, and laying out the design.
The architecture team begins by defining how the CPU is supposed to work. How many registers will it have? What’s the power budget? How many cores? How much cache? These and thousands of smaller details are all ironed out in the design document, which becomes the bible from which the final product is created.
Once the design is in place, it needs to be tested. How do you test a CPU that doesn’t exist yet? You run simulations. There are specific programming languages that chip designers use to build simulations of a CPU. Actual code is compiled and run on the simulated CPU, albeit much more slowly than on the final product. Those applications-code traces collected during the specification process are re-run on the simulation to make sure everything works as expected.
In the layout phase, the real process of building the CPU begins. Engineers use special software to route circuits into patterns that can then be processed in the lithography step. With high-performance PC processors, some elements of the logic layout are hand-tuned, while other aspects, such as cache line layout, may be automated. Chip companies often have prebuilt blocks in libraries that can just be dropped into the overall CPU layout.
Today’s processors also utilize multiple layers of semiconductors. Each layer needs to be laid out so that it can be connected to the others. The primary goal of the layout step is to create circuit patterns that are efficient yet simple enough that they can be manufactured. The first draft of the design undergoes verification, which runs more virtual tests on the layout to make sure connections are correctly made and circuits completed. The final layout is known as tape out, where the layout is compiled into an industry standard format and sent to manufacturing.
Note that these design-phase steps aren’t linear. Simulations, for example, will be run constantly, up until the first working silicon returns from the fab. Design is an iterative process, continuing to the point when the first chips come off the assembly line.
Here’s where we get into the physical processes of building our CPU. First, ultra-pure wafers of silicon are coated with the conductive material that will make up the final circuitry. Then the chip is baked at temperatures above 200 degrees C to remove any water or volatile contaminants.
Building a chip is essentially a photographic process. Photoresist—material that is light sensitive—is applied uniformly to the wafer, usually by spraying it onto the wafer while it’s spinning at high speed. The layer must be thin and very uniform. Once applied, the chip is again baked to dry the photoresist and make it more uniform.
The lithography step marks the chip’s design on the wafer by exposing the photoresist to light of specific frequencies. These intense beams of light, which shine through masks, define the layout of the circuits on the chip. Note that these beams are very narrow, so either the beam scans across the wafer, or the wafer is moved slightly (stepped) under the light beam. Today’s modern process technologies often use a hybrid of the scanning and stepping techniques. Another bake cycle removes imperfections left over from the lithography process.
The develop step removes the exposed photoresist, leaving behind patterns of circuits. Now the wafer has a layer of material with narrow “channels” laid out in the pattern of the CPU circuitry. But these patterns are not yet circuits. Next, chemicals are applied to the wafer that permanently remove the now exposed conductive material, which was initially coated on the chip in the wafer prep phase. The photoresist still on the chip resists the etching process, so only the circuit patterns are implanted into the wafer substrate.
The final step in the actual chip making process is stripping the remaining photoresist from the wafer surface. What’s left are many dies on the wafer, cleaned and ready to be processed.
Next, the entire wafer is tested to ensure it meets quality standards. The dies are then cut and sent to the packaging line, where the different layers are assembled into the chip packages we’re all familiar with. During the packaging process, function and validation tests are performed, which allow the manufacturer to sort according to clock speed and functional bins. This is where a Core 2 Quad Q9650 may be differentiated from a lower-clocked Q9550, for example.
Of course, this is a simplified overview of the process for building a modern CPU. You can find more details at websites including entries on Wikipedia for photolithography, photoresist, wafer creation, and more. One fairly technical, but still understandable overview of the lithography process can be found at Lithoguru ( www.lithoguru.com/scientist/lithobasics.html ).