AMD made its big comeback back in 2017 on the back of its Ryzen CPUs, which are still some of the best you can buy today, and it was all possible thanks to the company's brand-new Zen architecture. The success of Zen turned AMD from nearly broke to one of the most preeminent tech companies in the world, all within the span of six years. This is the story of Zen, how it saved AMD, and what the future of Zen might look like.

A brief history of Zen

The AMD Ryzen logo.
Source: AMD

In the late 2000s, AMD was down on its luck. Only a few years prior, its legendary Athlon desktop and Opteron server CPUs seemed poised to topple Intel, but eventually, AMD lost its grip and Intel cleaned up its act. AMD's Phenom CPUs just weren't cutting it against Intel's Core architecture, and something needed to change if AMD wanted to have a shot at leadership again. So, the company decided to develop this architecture called Bulldozer and bet that multi-threaded workloads were the future of computing.

Bulldozer wasn't just bad, it was objectively the worst thing AMD ever came up with. Its single-threaded performance was trash (first-generation FX chips were actually slower than the Phenom II CPUs they replaced), it consumed tons of power, and at the end of the day, its multi-threaded performance was at best mediocre. For the next six years, AMD would have to subsist on this awful architecture while Intel reached the peak of its supremacy.

Almost immediately after the Bulldozer debacle, AMD realized a simple rework wouldn't cut it and started working on a brand-new architecture. This architecture would be modeled after Intel's: high single-threaded performance, industry-typical cores and threads, and the kind of flexibility that made it suitable for everything from the lowest-end consumer CPUs to the highest-end server chips. AMD later named this architecture Zen, and the launch of its first Zen CPUs in 2017 marked a new beginning for AMD, and although Zen couldn't quite compare to Intel's Core architecture, it wasn't far off.

While the computing industry, CPU enthusiasts, and even AMD itself expected the road to performance leadership to be long, it was actually quite short. Zen 2, the successor to Zen, launched in 2019 and shocked pretty much everyone by blowing Intel out of the water. AMD racked up a massive lead in multi-threaded performance in pretty much every segment, had significantly better power efficiency in virtually every workload, and even surpassed Intel in single-threaded performance, which AMD hadn't been able to do for over a decade.

From here, the road just got easier for AMD. The server market was (and still is) the most important area for AMD to make progress in, and by the time Zen 3 came out in 2020, AMD controlled 7% of the market, up from nearly 0% before Zen came out. This was made all the easier thanks to how Intel absolutely screwed up its plans to launch powerful 10nm CPUs, leaving AMD to face off against outdated and practically obsolete 14nm chips, which are some of the worst Intel has ever made.

A render of a Ryzen 7000 CPU.

By the end of 2021 though, Intel finally got its act together and launched its 10nm Alder Lake chips. It became pretty clear that AMD lost track of the market and got too caught up in its performance leadership, as Intel had no competition below the $300 mark on the desktop since AMD never bothered launching budget Ryzen 5000 chips until Intel forced the issue. The months following the launch of Alder Lake were a bit rough for AMD, but it still held the upper hand in the server market and retook the gaming lead thanks to the Ryzen 7 5800X3D and its 3D V-Cache.

Today, Zen is on its fourth major iteration, with Zen 4 having launched in late 2022 with the Ryzen 7000 series and Epyc 4th-generation. This latest version of the Zen architecture is focused on high performance, which is in stark contrast to the original Zen architecture, which focused on better value. Although Zen 4 is significantly different than the original Zen, there are some fundamentals AMD hasn't let go of yet and probably won't be for some time.

CCXs, chiplets, and cores

A render of a Ryzen 7000 CPU with 3D V-Cache and two CCDs.
Source: AMD

While AMD has over the years improved many things in its Zen architecture, there are lots of things about Zen that have been fundamentally true since the very beginning, and a few new things that will shape Zen going forward. I'm talking about CCXs, chiplets, and cores, the fundamental aspects of modern Zen chips.

The Zen architecture is powerful, but it's not quite as flexible as competing designs from companies like Intel. While the smallest building block in most CPUs is the core, for Zen it's the Core Complex, or CCX. A CCX is a cluster of cores and can contain (at the time of writing) two, four, or eight cores, has its own L3 cache, and works with other CCXs in the same CPU. A CCX is essentially a full CPU unto itself, which is both a good and a bad thing. Each CCX is very capable on its own, but communication between CCXs takes a significant amount of time, which reduces performance.

For AMD, the generalized nature of the CCX makes it challenging to offer certain core counts. For example, if AMD wants to make a six-core CPU, it can't just develop a chip with six cores, because AMD doesn't have a six-core CCX. Initially, AMD only had the four-core CCX, so it needed to take a chip with two of those CCXs and disable a core on each to get a six-core CPU. Today, AMD takes a chip with an eight-core CCX and disables two cores on that to get down to six. Technically AMD can combine CCXs of differing sizes to get more options, but I'll discuss that later.

With Zen 2, AMD developed chiplets to make Zen even more powerful. While the original Zen architecture simply stitched together multiple CPUs to achieve higher core counts, Zen 2 chiplets introduced a radical concept by putting the CPU cores on its own chips and everything else on another. Chiplet design stands in opposition to traditional monolithic design, in which all CPU functions exist on a single chip. The chiplets with the cores are called Core Complex Dies (or CCDs), which can contain either one or two CCXs, and the chiplets with everything else are the I/O Dies (or IODs).

A render of an AMD Epyc Milan-X server CPU.
Source: AMD

There are many benefits with chiplets that align with AMD's goal of building CPUs thriftily. Firstly, it's cheaper to make lots of small chips as opposed to a big one with the same characteristics. Secondly, it makes it easy to make CPUs with super high core counts since all you have to do is add more chips. Perhaps the biggest advantage is flexibility, since AMD is able to cover pretty much the entire desktop and server market with one kind of CCD and two kinds of IODs. AMD also now has cache chiplets called 3D V-Cache for even more flexibility and customization.

AMD's latest innovation is the introduction of denser variants of Zen cores with Zen 4c. These dense versions of the Zen architecture are completely identical to the regular versions except are much smaller, allowing AMD's 16-core Zen 4c CCD to be the same size as the eight-core Zen 4 CCD. However, that increased density prevents c-type cores from hitting the clock speeds that regular cores can. This makes Zen c-variant cores more preferred for high-core count CPUs that don't need great single-threaded performance.

These kinds of cores are useful for consumer applications too. AMD's Phoenix 2 APU combines a two-core Zen 4 CCX with a four-core Zen 4c CCX, the first to combine CCXs of differing sizes. Using two different cores is called hybrid architecture, and the whole idea is that the regular cores are used for single-threaded workloads while the c-type cores help out in multi-threaded workloads. While this chip looks unusually specialized for AMD, it can actually also be used for lower-end Ryzen APUs in case the non-hybrid Phoenix chip isn't available.

With the Zen architecture, AMD has been singularly focused on how to cover the market in the broadest way without wasting time and resources developing processors, which AMD can't afford to do owing to its relatively small size. Rather than treating each segment of the computing industry differently, AMD uses a generalized approach and only develops a few designs and individual chips to cover everything. Whereas Intel made four designs for Alder Lake, which only covered desktops and laptops, AMD had a single Zen 3 CCX design used for desktop, laptop, and server CPUs.

The future of Zen

Being such an innovative and clever company, it's never easy to guess what AMD will do next. AMD has disclosed its plans to launch Zen 5 CPUs in 2024, but beyond that we don't know anything for certain. Perhaps we'll see AMD offer a wider spread of hybrid CPUs, perhaps even some that combine regular and c-variant CCDs to offer the best of both worlds for desktops and servers.

We also can't ignore AMD's competitors, principally Intel and Arm, when it comes to the future of Zen. While Zen is undeniably a good architecture, much of AMD's success since the introduction of the original Zen architecture is thanks to Intel's strategic mistakes throughout the 2010s. But not only has Intel finally mounted its own comeback, a new challenger is approaching as Arm is creeping into PCs and servers. If AMD wants to keep and improve its position, Zen will have to keep getting better every generation.