How are scientists able design processors with such large amounts of transistors? Do they do it with automation? How is it possible to do something like this at that scale?
In: Physics
It’s a machine made of lots of little parts. To get your head around it, think about the complexity of a mechanical watch. Lots of little cogs, springs, that all have a part to play in timing, moving hands, phase of the moon, whatever other features it has, lots of little parts, fitted together to create the whole.
Now think about an early computer pre transistor, a tube was used as the switch, it was big and there was only so much you could practically fit. The invention of the transistor with the discovery of the properties of semi conductors made that smaller, more complex machines could be made with smaller switches. Take that forward to the integrated circuit, you can now print the circuit, lots of little parts, all together making the whole machine.
The transistors on chips are not built on, they are built in so to say. They get chemically graftet directly into silicium with etching processes and particle injection. Thats how its possible to work on such a small scale. The Princinple stays mostly the same wether you build 100 or 100.000.000 transistors. The size is of course much trickier when it gets smaller. They use optics to project a large matrix onto much smaller size (not visible light though)
The design of CPUs and other integrated circuit are done by engineers not scientists..
First lest look at what a CPU die to contain. Let’s take a Ryzen 1000 series CPU that contains 4.8 billion transistors and 8 cores.
There is a lot of cache memory that is SRAM with 8 transistors per bit. There is 64kb instruction L1 +32kB Data L1 cache and 512kB L2 cache per core for a total of 606kB per core. The chip also has 16 MB L3 cha chase that is shard so the total is 8*606/1024+16= 20.7MB or cache, 1 byte is 8 bits so you need 8*8*20,7= 1324 = 1.3 billion transistors. This number does not include storage for the address of the data and other gates used so est say 1.5 billion transistors.
This is 1/3 of the total number of transistors of the CPU and the SRAM will just be copies side by side of the same design, Because of the repeating structure of RAM it is very space-efficient so it will not take 1/3 of the chip area but it still contains 1/3 of all transistors,
Then we have 8 cores and each of them is identical so design one and make a copy. So lest look at [an image of the chip](https://qph.fs.quoracdn.net/main-qimg-c31dcc2845979d508b5bd6e9ccd9d3cd) where you can see the 8 cores and the cache memory in between.
Let’s look at that single-core and the [and the general functional design](https://cdn.wccftech.com/wp-content/uploads/2016/08/AMD-Zen-CPU-Architecture-7.png) The important part here is that you see multiple schedulers, ALU, AGU, ADD, and MUL blocks. So a core in itself contains copies of the same functional block. Even inside them, the is lost of preparing the structure, you can make a 64-bit integer adder by making a full adder that adds bits and has 64 identical side by side,
Around the core and cache, it is transistors for power management and to communicate with the rest of the computer so you find memory controllers, PCIe interfaces etc. The memory controllers are the long structure to the bottom left and you see two identical but mirrored parts because you have two memory channels. If you look around you see a lot of repeating patterns and that is not by chance. For example, 24 PICe lanes out will result in a lot of identical parts.
So a CPU contains a lot of copies of the same thing that you need to design once.
If you look at how the part you need to design is done you will find that there are fundamental logical gates and other structures that need to be made with individual transistors but then when you make the more complex function you can reuse the design on the logical gate level, So just like on the large scale you make fundamental building blocks and reuse them to do what you like
The result is often that humans design the logical gates with aid of software to make them fast and energy-efficient. You then create the function on a higher level with code in a programming language that is relatively similar to the programming language that the software on a computer. Look at [Verilog](https://en.wikipedia.org/wiki/Verilog) which is a hardware description language, if you have done any programming the code example is quite understandable and that is code that a compiler can use appreciated gates to create a layout that can be made in silicon.
It is on this code level the function of the CPU is tested. So you have a software description of the CPU that you can use it a test so it functions correctly.
You then let the software create what you wrote in the high-level code with the low-level structures that is designed, it can also lay them out on a die, So relatively simple test code can result in lots of transistors.
Depending on what part you look at there will be more or less human interaction. You spend time making the parts that limit the speed of the CPU and try to optimize them to keep the speed up, you den might make a completely custom design. But there will be another part that the computer-generated variants are good enough
So chip design is copies of the same part both at the top level and at the lowest level where the middle part is made in code that a computer makes of the lower level part or customer design if required for speed, efficiency etc.
A lot of the process is automated and there is a lot of repetition. There is not someone sitting there determining all of the individual transistors needed and where they go. The processor is designed using a block diagram and then modeled using a hardware description language which translates it to a lower level gate design and low level transitor layout.
Additional programs are used to create laser etched glass masks for the lithography part of the fabrication process.
There are two related questions here:
1) How do they design things with millions of transistors?
2) How do they manufacture things with at that scale?
Answers:
1) Hierarchy and Automation.
When a designer designs a processor, they are not choosing to place individual transistors, they are placing more complex devices that (generally) have already been designed. The most basics of these are logic gates, which are collections of transistors that map onto the boolean logic functions: AND, OR and NOT. From these gates, you can build more complex devices: multiplexers, encoders & decoders, full and half adders, registers etc. The designers job is generally to decided which versions of these devices are needs, and how they are to be connected in order to make the processors match the desired specifications (power, space, speed and cost).
Once the designer is finished, the automation takes over to do the layout. Layout is actually a hard problem (NP-Hard to be specific but not ELI5), as the individual connections between registers need to as short as possible but not overlap. In most cases, there will be simulations done before the design is manufacturer and the design may manually adjust components with CAD tools to get the performance they desire. The same CAD tools would be used to design the standard devices used throughout the hierarchy.
2) Photolithography.
First, a bit about how transistors work. Transistors work because it’s it relative easy to make “doped” silicons, which makes it easier or harder for electrons to pass through it. By alternating between the two kinds of doped silicons, you can create or close a channel between them by applying an electromagnetic field. This allows them to operate as voltage controlled switches. You can get these effects with extreme small layers of silicons, at the scale of 10s of molecules thick. At that scale, UV light can be used to crave and etch away shapes in a layer. So the basic process goes like this: A layer of silicons or metal connector is applied to the wafer, then UV light is used to etch away the parts they don’t need, and this is repeated countless times until all of the required layers are created, then a protective layer is applied. However, this isn’t done for individual transistors that are joined together at a later point. Instead, an entire device is built together, at once on a chip, and they are small enough that a multiple chips at built in at once on a single wafer.
They started small and worked their way up. The first integrated circuits (chips) had only a few transistors on them, but they quickly got better at design and manufacturing multiple transistors on a single chip, to the point where many microchips have billions of transistors.
Early microchips were designed by hand. This was done from the earliest experiements in the 1950s until the late 1970s, by which point there were thousands or even tens of thousands on transistors on a chip. The blueprints for the chip became so huge that the designers printed them out the size of a tennis court, then crawled around on them with pens and pencils to make changes.
Around this time, the first automated chip layout software became available. This software allowed a computer to do a lot of the detail work of routing individual wire traces, while the human designer made higher level decisions. This software has gotten better and better over time. In large part, the capability of the automated design software has dictated how complex chips can be, as the software got better, the chips got bigger.
Latest Answers