After more than a decade of dominance, Intel gave way to its main competitor, AMD. Jeremy Laird tried to figure out where Intel got on the wrong path and how it would respond to the enemy.
So what exactly happened to Intel? The once undisputed leader in the production of processors and microcircuits is now inferior to rivals in almost all possible indicators. AMD CPUs turned out to be more sophisticated, and TSMC production technology was more efficient. It looks like Intel has completely lost its way.
Even in the mobile PC market, where the manufacturer has been the absolute leader for decades, Intel processors have given way to AMD's new Renoir hybrid.
Things are so bad that Apple has announced plans to cut ties with the manufacturer and start producing its own ARM-based chips. Worse, rumor has it that Intel itself is considering a partnership with TSMC to release certain products in the future, including the first custom graphics card. In fact, this can be a complete humiliation for the corporation.
Or is this just a guess? Despite all the difficulties, Intel reached a record $ 72 billion in revenue last year. In reality, the manufacturer's biggest problem is that it is not keeping up with the dynamics of demand from so-called hyperscale data centers. These are companies like Amazon, Microsoft, Google, Facebook and others that simply lack enough Xeon processors. Meanwhile, there are good reasons to believe that Intel will soon return to its previous rut โโregarding the production of chips and CPU microarchitectures.
How can you describe the troubles and misfortunes of Intel in a nutshell? โ10 nanometers,โ I would say. And it's not just the failure of chip technology - such arguments can be made in favor of any microarchitecture manufacturing company that has been resting on its laurels for decades. But 10 nanometers! This is a disaster.
This term โ10 nanometersโ refers to a process or assembly used to manufacture computer chips. 10nm in theory is the size of the smallest components inside a chip. In practice, however, process names and actual sizes of components, such as transistor gates inside a desktop processor, have ceased to relate recently. And, most likely, there is no such component inside the Intel processor that is really 10 nm in size.
This lack of direct relationship between component size and assembly description becomes more problematic when it comes to comparing competing manufacturers' workflows. But more on that later. And now we are interested in Intel's 10nm process technology and its disadvantages. It was originally expected to be rolled out back in 2015. It is now the second half of 2020, but the range of products with 10nm chips is small. You will not be able to buy desktop PCs or server CPUs built on the above technical process. Only mobile processors for laptops and tablets have switched to 10nm technology, and even then only those with low and ultra-low power consumption. The rest have been updated to 14 nm.
These facts must be considered taking into account the standard adopted by Intel itself - Moore's Law, and therefore it is necessary to take into account the resistance of the laws of physics that chip designers have faced in recent years. However, even greater difficulties in semiconductor manufacturing can arise when individual transistors reach the size of a handful of atoms and undergo mysterious quantum effects, such as tunneling. But that's a completely different story.
Most likely, all of Intel's problems boil down to excessive ambition, obsolescence of a particular production technology, perhaps complacency and lack of investment.
According to Intel chief executive Bob Swan, Intel's 10nm problems are โkind of a derivative of what we have done in the past. Then we tried to win no matter what. And when times were particularly hard, we set even more ambitious goals. And that's why it took us more time to achieve them โ.
High expectations from microcircuits
For a 10nm technology node, this ambitious target translates into a 2.7x increase in transistor density. In other words, a 10nm node contains 2.7 times more transistors per unit die area than a 14nm node. More specifically, the 14nm processors contain 37.5 million transistors per square millimeter, while one square millimeter of 10nm crystals contains 100 million transistors. The dramatic increase in transistor density makes the 10nm technology much more ambitious compared to other process technology.
The 2.5-fold increase in density and the transition from 22nm to 14nm technology was impressive, however, the transition from 32nm to 22nm represented a 2.1-fold increase in density and a 2-fold increase in density from 45nm to 32nm. 3 times. Understanding these changes helps explain the differences between Intel and competing nodes. For example, Intel's 10nm technology implies a density of 100.8 million transistors per square millimeter. This figure is slightly higher than TSMC's figure of 96.5 million transistors (later TSMC announced 113.9 million transistors per square millimeter for an improved 7-nm process technology). All three 7nm nodes from Samsung also fall short of the 100M mark.
This is because Intel's 10nm technology was very ambitious - so much so that in 2017 the company added the "Hyper Scaling" label to draw attention to the increased density. In hindsight, it can be argued that expectations were too high. This is because Intel made an end node based on current far-UV (DUV) lithography. In a nutshell, the size of the components in a microcircuit is determined by the wavelength of light used in lithographic processes. These processes engrave the components on the surface of the silicon substrate, and PC processors are cut from the silicon wafers.
This is not a direct relationship. Various techniques and ancillary options can also have an impact, such as masks actually used as a multiplier that reduce the size of components below the actual wavelength of light.
DUV chip making equipment uses UV light with a wavelength of 193 nm. However, there is a limit to the density of transistors at a given wavelength. Intel has exceeded this limit.
The result is a shameful five-year delay in product release. It's eternity in terms of Intel volume dynamics and Moore's Law. Even now, there are indications that the 10nm process is not what it should be. So Ice Lake, the new tenth generation mobile processors, accelerate slower than their 14nm predecessors. The fastest 10nm Ice Lake processors, the Core i7-1065G7, reach top speed at 3.9 GHz, while the 8th Gen Core i7-8665U is a whopping 900 MHz faster. That's a hell of a lot, which means something's going wrong in the production process.
Another proof that the 10nm process has fallen short of Intel's expectations is the twinning of low-power 10th generation processors. Along with the current Ice Lake CPUs, a new Comet Lake family is being released, and both are classified as 10th generation.
Like Ice Lake, Comet Lake mobile processors come in low-power and ultra-low-power formats.
But unlike Ice Lake, Comet Lake uses 14nm, not 10nm, and extends to 6-core models at a maximum clock speed of 4.9GHz.
As a result, you can already buy a laptop with a processor that has the Intel 10th generation logo, but what is inside the box may differ from the declared one. If the processor is 2 or 4-core, it can be low-power or ultra-low-power. And also 10-nanometer or 14-nanometer. It can be based on the 2015 Skylake microarchitecture or the completely new Sunny Cove, and also be considered an Ice lake.
Microarchitecture problems
The mention of Sunny Cove naturally brings us to another big Intel failure - microarchitecture. Until the release of 10nm Ice Lake chips for ultraportable laptops late last year, a huge number of processors for desktops, laptops and servers were based on the 14nm process technology, which debuted in 2014, and the Skylake architecture, which appeared in 2015. Both have been ruled thousands of times, but there have been no major changes in the updates.
Moreover, since the introduction of the Nehalem microarchitecture in 2008, Intel could only offer 4 processor cores for popular PC models. This continued until the 2017 release of the Coffee Lake microarchitecture, an evolved version of Skylake, and the subsequent increase to six cores. For nearly a decade, Intel hasn't increased the number of cores for its mainstream product models.
In a little less than two and a half years, Intel raised the bar to 10 cores for popular desktop processors with the release of Comet Lake, a Skylake rebuild of the 14nm CPU family. It turns out that for 10 years there were no shifts, and then there was an increase of 2.5 times in a short period of time. What could have led to such a sharp increase in the number of nuclei after a long stagnation? The reason for this is the emergence of the Zen architecture from AMD and the Ryzen processors, the first generation of which was released in 2017. Simply put, Intel was lazy until it got a competitor.
Of course, even with ten cores, Intel is far behind AMD, which currently offers 16 cores in popular PCs with 3rd Gen Ryzen processors. Their advantage also lies in the fact that they are based on TSMC's 7nm process technology.
In the mobile segment, Intel is no better. AMD's new line of 7nm Renoir APUs have eight Zen 2 cores with 15 watts. Intel only managed to make a 6-core Comet Lake Core-i7 10810U as a competitor. This is a processor with a clock speed of only 1.1 GHz. The 15-watt Ryzen 7 4800U is packed with 8 cores and clocked at 1.8 GHz. Not a flattering comparison.
A look into the future
Here's the prosecution's version. The past few years have not been technologically fruitful for Intel. George Davis, the company's CFO, said of the 10nm flop: โThis technology node is definitely not going to be the best in Intel's history. It is less efficient than the 14nm process technology and less efficient than the 22nm process technology. " But are the consequences of Intel's current difficulties really so disastrous?
From a financial perspective, this question can be answered unequivocally - no. In reality, not only is the current situation not so bad, in fact there is no problem at all. Intel's revenues hit record highs in 2019. Since mid-2018, its sales have not fallen due to technological stagnation, and the manufacturer itself has experienced difficulties in meeting the demand for its 14nm processors.
If you dig deeper, you can come to the conclusion that at least part of the problem lies in the process. The number of cores in Intel server processors has skyrocketed with the advent of the 14nm era. Intel now offers as many as 28 cores in a single processor die. This means that the more cores in the same process, the fewer processors can be extracted from one semiconductor wafer, which in turn can lead to a limited supply.
But whatever one may say, Intel is not experiencing any financial difficulties, and this circumstance is the main reason why the manufacturer can give a decent answer to competitors in terms of products and technology.
And this effect is already visible. Ice Lake processors have a new microarchitecture known as Sunny Cove. It improves performance per clock by 18% over Coffee Lake, a refinement of the Skylake microarchitecture.
And this is just the beginning. A decisive factor in the revival of Intel's microarchitecture was the inclusion of Jim Keller on the team, who led the microprocessor development team.
Although he plans to leave this post in six months, the contribution he can make to the development of the company cannot be underestimated. Keller is one of the most respected, if not the most respected microprocessor architect.
It became famous for the development of the microarchitecture of the K8 processor, codenamed Athlon 64, and the first chip from AMD to compete with Intel. Later, Keller worked at Apple, designing a series of ARM-based processors of its own production, which later took the leading positions in the smartphone and tablet market. In 2012, Keller returned to AMD, leading the development of the Zen microarchitecture and once again providing AMD with the tools to fight Intel. After a brief tenure as head of Tesla's electric vehicle development team, Keller took over as senior vice president of Intel in April 2018.
Given the lapse of time between the design and microarchitecture of the processors and the launch of the product launch, it is highly unlikely that the new Sunny Cove cores inside Ice Lake processors are Keller's work. The same will be true for the Willow Cove architecture that follows Sunny Cove. It is planned to be released at the end of this year for a family of 14-nm backported processors, that is, using the "reverse transfer" of the new microarchitecture to the "old" technical process, Rocket Lake processors.
The Golden Cove microarchitecture will take an even bigger step forward and will lay the groundwork for planned Alder Lake processors later next year. But even Golden Cove cannot be considered a complete creation of Keller. To do this, we need to wait until Ocean Cove comes out in 2022 or 2023, although Keller's imminent departure will mean that his influence on the project will likely be somewhat limited.
There is no official information about Ocean Cove yet. Recently there have been rumors that the performance of this microarchitecture will be 80% higher than Skylake. While these are just rumors, we know for sure that Keller has an outstanding track record and that Intel has an ambitious strategic plan that goes far beyond what it did years ago. As Keller said, "We plan to increase the number of transistors by 50 times and do everything to squeeze the most out of each stack."
At the same time, 7nm CPUs following the problematic 10nm processors will not face the same constraints as their predecessors. For the production of 7-nm processors, lithography of the extreme ultraviolet range (EUV) with a wavelength of up to 13.5 nm will be used. In other words, the 7nm process technology has changed dramatically. Time will tell, but now we can definitely say that Intel's forecasts are too optimistic.
Intel plans to accelerate the transition from 7nm to 5nm and beyond. This means that the manufacturer will actively develop new technology as opposed to the current costly one, even if it requires investment in research and development. Moreover, with the involvement of EUV lithography, Intel expects to return to the previous rate of production - once every 2 years, starting with a 7-nm process technology at the end of 2021 and reaching the release of 1.4-nm technology in 2029. โI think EUV will help us get back to the pace at which Moore's Law transistors are increasing,โ Davis said.
All this taken together gives the impression that Intel is returning the standards for creating the most advanced architectures and the fastest processors. Whether this will happen is another question. AMD is now arguably in a better position than Intel, despite the latter putting in significantly more effort. AMD's strategic roadmap for microarchitectures, including Zen 3 and Zen 4, coupled with TSMC's technology solutions, will foster competition between the two manufacturers. However, we will not predict Intel's defeat.
After all, the last time NetBurst and Pentium 4 made their way and Intel stalled, the answer was the Core dynasty and the leadership in the processor market for a decade.