If you’ve ever looked at the spec sheet or an advertisement for a CPU, GPU, or even a fully built device like a laptop or desktop, you’ve probably seen some hype about how it uses a 7nm or 5nm or even 4nm process. node or process node. But like many technical specs, the process node is much more complicated than a simple number, rarely explained by marketing, and not something you should really care too much about. Here’s everything you need to know about process nodes, what they actually mean for computer chips.
Process Nodes: A major reason why processors get faster every year without a hitch
Source: XDA Developers
Process nodes are all about chip manufacturing, also known as fabrication or “fabbing,” which takes place in factories called fabs or foundries. While virtually all chips are made using silicon, there are several manufacturing processes that foundries can use, which is where the term process comes from. Processors are made up of many transistors, and the more transistors the better, but since chips can only be so big, putting more transistors in a chip by reducing the space between transistors to increase density is a big deal. The invention of newer and better processes or nodes is the main way to achieve greater density.
Different processes or nodes are distinguished by a length historically measured in micrometers and nanometers, and the lower the number, the better the process (think rules of wave). This number used to refer to the physical size of a transistor, which manufacturers want to shrink when creating a new process, but after the 28nm node, this number became arbitrary. TSMC’s 5nm node isn’t really 5nm, TSMC just want you to know it’s better than 7nm and not as good as 3nm. For the same reason, that figure cannot be used to compare modern processes; TSMC’s 5nm is completely different from Samsung’s 5nm, and even in the case of TSMC’s N4 process, it is considered part of TSMC’s 5nm family. Confusing, I know.
However, new processes not only increase density, they also tend to increase clock speed and efficiency. For example, TSCM’s 5nm node (used in Ryzen 7000 and RX 7000 processors) compared to the older 7nm process can deliver either 15% higher clock speed at the same power or 30% lower power at the same frequency, or a combination of both on a sliding scale. However, frequency and efficiency gains were much more dramatic until the mid-2000s, as shrinking transistors directly reduced power consumption in older processes, a trend called Dennard scaling.
The death of Moore’s law and what process nodes are involved with it
The main motivation for companies to adopt newer processes is to keep up with Moore’s Law, an observation made in 1965 by legendary semiconductor figure Gordon Moore. The original law stated that the growth rate of transistors in the fastest CPU doubles every two years; if the fastest processor has 500 million transistors in one year, in two years there should be one with a billion transistors. For more than 40 years, the industry was able to keep up with this pace by inventing new processes, each with a higher density than the last.
In the 2000s, however, the industry began to run into problems. First, the Dennard scaling collapsed around 65nm to 45nm in the mid-2000s, but after the 32nm process came out in the late 2000s and early 2010s, all hell broke loose. For most foundries, this was the last major hub they would supply for many years. TSCM’s 2014 20nm was just bad and only the 16nm process in 2015 was a worthwhile upgrade from 28nm in 2011, Samsung didn’t get to 14nm until 2015 and GlobalFoundries (split from AMD’s fabs in the 2000s) had to lease Samsung’s 14nm instead of to make it yourself.
A notable exception to this turmoil was Intel, which successfully got its 22nm process out the door in 2011. However, Intel’s release schedule and process quality started to drop after the 22nm mark. The 14nm process was supposed to be released in 2013, but was released in 2014 with low clock speeds and many defects. Intel’s ridiculous goals with its 10nm node ultimately doomed it to development hell, missing its 2015 launch period. The first 10nm chip arrived in 2018 and it’s one of Intel’s worst CPUs ever. Intel’s 10nm, renamed Intel 7 for marketing purposes, wasn’t fully ready until 2021.
The latest disaster concerns TSMC’s 3nm node, which provides a significant density improvement in logic transistors (which include the cores of CPUs and GPUs, among others), but literally no improvement at all in cache density, otherwise known as SRAM. Not being able to shrink the cache is a total disaster, and foundries may run into similar problems on future nodes. Even if TSMC is the only fab struggling to shrink the cache, it is also the largest chip producer in the world.
When you read about the death of Moore’s law, it means this, because if companies can’t increase density year after year, then transistor count can’t increase. If the transistor count can’t go up, that means Moore’s law is dead. Today, companies are focused on tracking the performance implications of Moore’s Law, rather than the technical implications. If performance doubles every two years, everything is fine. AMD and Intel are using chiplets to increase both transistor count and performance while reducing costs, and Nvidia relies solely on AI to take up the slack.
Ultimately, process nodes are just one factor that determines whether a chip is good
When you consider that a new process can make a chip smaller, clock faster, and more efficient, all without major design or architecture changes, it’s easy to see why processes are so important. However, other factors such as packaging (such as chiplets or tiles or stacking chips) and AI are increasingly becoming viable ways to add value to a processor by improving performance or adding features, not to mention simple optimization in software. The death of Moore’s Law isn’t ideal, but it isn’t the end of the semiconductor industry.
Furthermore, because nodes are named for marketing reasons, there’s no real reason to estimate a chip’s competence based solely on process; For example, Intel’s 10nm is about as good as TSMC’s 7nm, despite 7 being less than 10. However, it’s also true that a process isn’t the only function that matters in a processor. Plenty of CPUs, GPUs, and other processors sucked despite being on good nodes, like AMD’s Radeon VII, which was a full process node for Nvidia’s RTX 2080 Ti yet was so slow it was one of the worst GPUs ever.
By itself, the process node of a chip means nothing. It would be like buying a CPU based solely on how many cores it has, or a console because it has blast processing. What really matters in a processor is its actual performance, which boils down to other hardware specs and how well optimized applications are for that hardware. If you just want to know what the best CPU or GPU or laptop is, the process node won’t tell you. It just tells you who made the chip.