THE EMPIRE STRIKES BACK
After more than a decade of dominance, Intel has fallen behind its main rival, AMD. Jeremy Laird investigates where Intel went wrong and how it will fight back
WHAT, EXACTLY, has happened to Intel? The once unassailable leader in both PC processors and chip manufacturing generally now looks second best by almost every conceivable technological metric. AMD’s CPUs appear cleverer. TSMC’s manufacturing technology seems more advanced. Intel, it seems, has completely lost its way.
Even in the mobile PC space, which Intel has absolutely owned for decades, AMD’s new Renoir APU has the edge. Likewise, how is it that AMD has offered PCI Express 4.0 on desktop PCs for a year, but Intel hasn’t even announced future support for what is a crucial highspeed interconnect?
Things are so bad that Apple is said to be planning to ditch Intel in favor of its own ARM-based processors. Worse, Intel itself is rumored to be considering TSMC as a manufacturing partner for selected future products, including its first consumer graphics card in decades. Truly, that would be the ultimate humiliation.
Or would it? Last year, despite those issues, Intel enjoyed record revenues of $72 billion. In fact, in pure commercial terms Intel’s main problem has been keeping up with demand for its chips from so-called hyperscalers. That’s the likes of Amazon, Microsoft, Google, Facebook et al, who simply can’t get enough of Intel’s Xeon processors. Meanwhile, there are good reasons to think that Intel will soon be back on track when it comes to both chip manufacturing and CPU architectures.
WHAT SINGLE thing summarizes Intel’s woes? Try “10nm.” Chip production technology isn’t Intel’s only failing—a strong case could also be made for the company becoming ever more complacent over the past decade when it comes to CPU architectures—but 10nm? What a disaster.
By 10nm, of course, we mean a specific production process or node used to manufacture computer chips. The measurement 10nm refers, in theory, to the size of the smallest features inside the chip. In practice, the monikers attached to production processes and the actual size of components like transistor gates inside a PC processor have become somewhat estranged in recent years. There probably isn’t any individual feature that actually measures 10nm within a “10nm” Intel CPU.
That arguable lack of direct correlation between feature size and the description of a given node becomes even more problematic when it comes to comparing process technologies from competing manufacturers. But more on that in a moment. For now, what matters is that 10nm is Intel’s latest production process and surely its most troubled. Originally, 10nm was due to arrive way back in 2015. Here, in the second half of 2020, 10nm still only makes up a small minority of Intel’s production. You can’t buy a 10nm desktop or server CPU from Intel. Only mobile processors for laptops and tablets have moved to 10nm, and even then just some of Intel’s low and ultra-low-power range have moved to 10nm. Others have been refreshed on 14nm.
The context for all this—the selfimposed metric by which Intel can unambiguously be said to have fallen short—is Moore’s Law. For the uninitiated, Moore’s Law is the assumption that computer chips either double in complexity or halve in cost—or some mix of the two, depending on your objectives— every couple of years. It dates back to observations made between 1958 and 1965 by Intel co-founder Gordon Moore. It has since proven remarkably prescient. At least it did until the last decade, when the first signs emerged that chip engineers were beginning to bump up against the laws of physics.
But the wider trials and tribulations
of semiconductor manufacturing—as individual transistors approach the size of a handful of atoms and begin to suffer from arcane quantum effects such as tunnelling—are a story for another day and another feature. It’s the specifics of Intel’s 10nm process that matter for now.
Very likely, Intel’s problems boil down to a combination of over-ambition, the end of the line for a certain manufacturing technology, and, just maybe, complacency and a lack of investment. According to Intel CEO Bob Swan, Intel’s problems with 10nm are “somewhat a function of what we’ve been able to do in the past, which in essence was defying the odds. At a time when it was getting harder and harder, we set a more and more aggressive goal. From that, it just took us longer.”
CHIP AMBITIONS
For the 10nm node, that aggressive goal means improving transistor density by a factor of 2.7. In other words, for a given area of processor die area, the 10nm node contains 2.7 times as many transistors as the previous 14nm node. If it’s more numbers you want, the 14nm process clocks in at 37.5 million transistors per square millimeter, while the 10nm node achieves just over 100 million by the same metric. That dramatic advance in density makes 10nm demonstrably more aggressive than other recent node transitions from Intel. The 2.5 times step from 22nm to 14nm was quite impressive, but 32nm to 22nm represented a density improvement of 2.1 times, while 45nm to 32nm was pegged at 2.3 times.
Comprehension of these variable density improvements helps to explain the aforementioned difficulty comparing production nodes from competing manufacturers. By way of example, Intel’s 10nm density of 100.8 million transistors per square millimeter is actually slightly superior to the 96.5 million transistors achieved by rival manufacturer TSMC’s initial 7nm node (TSMC claims 113.9MTr/ mm2 for its refined 7nm node). All three of Samsung’s 7nm nodes also come in at slightly under the 100 million mark.
The point is that Intel’s 10nm was very ambitious—ambitious enough that in 2017 Intel applied the “Hyper Scaling” label to the increase in density it achieves. With the benefit of hindsight, that ambition now looks like overreach. That’s because Intel made 10nm its final process node based on existing deep ultraviolet lithography (DUV) manufacturing technologies. To cut a long story short, the size of features in a chip are a function of the wavelength of light used in the lithography processes that etch features into the surface of silicon wafers, from which the likes of PC processors are eventually cut.
It’s not a direct relationship. Various techniques and tweaks such as masks are used, in effect, to act as multipliers, reducing feature sizes beyond the actual wavelength of the light in question. Chip production facilities, or fabs, with the most advanced DUV technology, including Intel’s 10nm products, use ultraviolet light with a wavelength of 193nm. Even so, there is a limit to the transistor density that can be achieved with a given wavelength. Intel was pushing up against that limit with 10nm using DUV technology.
Intel has upped the ante to 10 cores for mainstream desktop processors
The result, infamously, has been a delay of as much as five years, a relative aeon in the context of Intel’s track record and Moore’s Law. Even now, there are signs that 10nm isn’t what it ought to be. The new 10th-generation Ice Lake mobile processors, for instance, actually clock lower than their 14nm predecessors. The fastest 10nm Ice Lake Core i7-1065G7 CPU tops out at 3.9GHz. The 8th-gen Core i7-8665U is fully 900MHz faster. That’s a heck of a gap, and it’s symptomatic of a production process that isn’t quite right.
Another indicator that 10nm still isn’t what Intel would really have wished for is the peculiar doubling up of 10th-gen lowpower CPUs. Alongside those Ice Lake chips is the new Comet Lake family, also dubbed 10th generation. Like Ice Lake, mobile Comet Lake comes in low and ultra-low-power formats. But unlike Ice Lake, it is a 14nm family, not 10nm, and extends to six-core models and up to a peak clockspeed of 4.9GHz.
The upshot of which is that, right now, you can buy a laptop with a CPU branded with Intel’s latest 10th-gen logo. But what’s inside can vary dramatically. If it’s dual or quad-core, it may be low or ultra-low power. It might be 14nm or 10nm. It could be based on the Skylake architecture from 2015, or a brand new architecture with the latest Sunny Cove CPU cores, as seen in Ice lake.
ARCHITECTURAL WOES
Speaking of Sunny Cove, it brings us rather neatly to Intel’s other major failing beyond manufacturing, namely architectures. Up until the release of 10th-generation Ice Lake CPUs for ultra-portable laptops late last year, Intel’s entire CPU product stack for desktops, laptops, and servers was based on the 14nm process that debuted in 2014, and the Skylake architecture of 2015. Both have been revised on numerous occasions, but the updates were just that—revisions, not step changes.
What’s more, since at least 2008 and the Nehalem architecture, Intel has offered a maximum of four cores for mainstream desktop processors. You could argue that quad-core for mainstream desktops even predates Nehalem. But at the very least, Intel’s mainstream offerings remained quad-core at best from 2008 until the release of the Coffee Lake update to Skylake in 2017, and with it an increase to six cores. For nearly a decade, Intel didn’t increase its mainstream core count.
Fast-forward a little over two and a half years, and Intel has upped the ante to 10 cores for mainstream desktop processors with Comet Lake, yet another respin of Skylake and yet another family of 14nm processors. So, that’s nothing for nearly 10 years, then a 2.5 times increase in the space of just a few. The reason for this dramatic increase in core count after such sustained stagnation? That’ll be AMD’s Zen architecture and Ryzen CPUs, the first generation of which arrived in early 2017. Put simply, Intel got somewhat lazy during the pre-Ryzen period, when AMD offered limited competition.
Even at 10 cores, of course, Intel is still a long way behind the 16 cores AMD currently offers in the mainstream desktop space with its 3rd-generation Ryzen processors, which have the benefit of TSMC’s cutting-edge 7nm process. In the mobile space, it doesn’t look much better for Intel. AMD’s latest 7nm Renoir APU brings eight Zen 2 cores into the lowpower 15W segment. Intel can only muster a six-core Comet Lake Core-i7 10810U to compete, a processor with a mere 1.1GHz baseclock. The 15W Ryzen 7 4800U packs eight cores and a 1.8GHz baseclock. It’s not a flattering comparison.
LOOKING TO THE FUTURE
So that’s the case for the prosecution. It hasn’t been a great few years for Intel from a technology perspective. As Intel’s CFO, George Davis, said of the 10nm debacle: “This just isn’t going to be the best node that Intel has ever had. It’s going to be less productive than 14nm, less productive than 22nm.” But is Intel’s predicament really that disastrous?
In pure financial terms, the resounding answer is no. In fact, not only is the situation not that bad, there’s no problem at all. 2019 was a record year for Intel revenues. Since the middle of 2018, far from sales falling off due to technological stagnation, Intel has been struggling to keep up with demand for its 14nm chips.
Dig a little further into the details, and it seems likely that at least part of the problem is still technological. Intel’s server chips have increased in core count dramatically during the 14nm era. Intel now offers as many as 28 cores in a single processor die. More cores on the same process means fewer CPUs can be extracted from a single wafer, which in turn can constrain supply.
But however you slice it, Intel has not been suffering financially, which is the most important reason why you can expect Intel to respond in terms of product and technology. As for resources, Intel retains a huge advantage over its only rival in the PC processor market, AMD.
That response has already started to show. Ice Lake debuted Intel’s first new CPU core architecture since Skylake, known as Sunny Cove. It delivers performance-per-clock improvements of around 18 percent over the Coffee Lake revision of the Skylake architecture.
But that’s just the beginning. Central to Intel’s architectural resurgence was the addition of Jim Keller, heading up Intel’s CPU development team. Although he’s departing in the next six months, it’s hard to disregard what his experience may bring. Keller is one of, if not the most, highly regarded processor architects in the business. He made his name as lead architect of AMD’s K8 technology, the chip otherwise known as the Athlon 64, and the first CPU from AMD that really took the fight to Intel. Keller later worked for Apple, kick-starting the series of in-house
ARM-based processor designs that have dominated the smartphone and tablet segments. In 2012, Keller returned to AMD, driving the development of Zen and once again giving AMD the tools to fight Intel. After a brief stint at EV maker Tesla, Keller arrived at Intel in April 2018.
Given the lead times between the initial design and conception of a CPU architecture and its inclusion in products available to buy, it’s very unlikely that the new Sunny Cove cores inside Ice Lake processors are Keller’s work. The same is likely true of the follow-up to Sunny Cove, known as Willow Cove, which may appear later this year in the Rocket Lake family of 14nm “backport” CPUs and is not thought to be a major advance on Sunny Cove.
Next up is Golden Cove, which will be a bigger step forward late next year when it forms the basis of Intel’s upcoming Alder Lake CPUs. But even Golden Cove probably isn’t what you might call a pure “Keller” processor. For that we’ll have to wait for 2022 or 2023 and Ocean Cove, although his impending departure from Intel will mean his influence on even this project will likely be somewhat reduced.
Almost nothing official is known about Ocean Cove. Recently, rumors emerged indicating that Ocean Cove could deliver as much as 80 percent more performance per clock than Skylake. For now, that is speculation. All we know for sure is that Keller has an outstanding track record and that Intel’s roadmap is more aggressive now than it has been for many years. As Keller himself has said, “We have a roadmap to 50 times more transistors and huge steps to make on every single piece of the stack.”
Meanwhile, the follow-up to the troubled 10nm node, Intel’s upcoming 7nm process, doesn’t face the same limitations as 10nm. For 7nm, Intel is transitioning to extreme ultraviolet lithography (EUV) with wavelengths as small as 13.5nm. 7nm, in other words, is a fundamental reset.
Time will tell, but Intel is certainly making bullish predictions for 7nm and beyond.
Intel plans more overlap between its upcoming nodes, including 7nm, 5nm, and the processes that follow. That means pressing on with new technology as opposed to milking nodes for every last cent, even if that comes at an R&D cost. “In order to regain process leadership, we had to accelerate the overlap between 10nm and then 7nm and 5nm. Starting particularly in 2021, the performance of 10nm will intersect with investment in 7nm, and we’re also willing to start an investment in 5nm, all of those elements combined,” says Intel CFO George Davis.
In fact, Intel expects to return to a twoyear release cadence based on the new EUV era, starting with 7nm late next year and extending all the way to the release of the 1.4nm node in 2029. “I think EUV will give us back a degree of the traditional Moore’s Law cadence,” says Davis.
Taken together, it paints a picture of Intel reverting to the norm of creating the most advanced processor architectures and the world’s fastest CPUs. Whether that happens is another matter. AMD is arguably in a better position now than ever before to maintain the pressure, even if Intel significantly ups its game. AMD’s architecture roadmap, including the likes of Zen 3 and Zen 4, combined with TSMC’s manufacturing prowess, will make for stiff competition. But we’ll not be betting against Intel. After all, the last time Intel stumbled, back in the dying days of NetBurst and the Pentium 4, its response was the Core dynasty and a decade-long domination of the CPU market.
Next up is Golden Cove, which will be a bigger step forward late next year