Breakfast Nibbles 2021: Predictions for the Year, part 3

The first part of my predictions for 2021 was two days ago, and the second part was yesterday.

Process

Everything in the first two posts has been driven by markets, not the basic technology. But the semiconductor industry's dynamo is the process roadmap. There are just three leading-edge manufacturers remaining.

2021 should be the year that Intel gets its 10nm process into full production. Intel names its processes in a way that makes it seem behind the foundries. But their 10nm is pretty much equivalent to the foundries' 7nm if you look at the widths and spacings of the various layers. Its next-generation process, which it calls 7nm, is equivalent to the foundries' 5nm.

Intel may spin out its manufacturing business as a separate company, in much the same way that AMD did a decade ago to create GLOBALFOUNDRIES. Otherwise they face the classic fab pressure: If you have a fab, the mentality is always to "fill the fab".

The reality of a company mostly building other people's designs (ASIC or foundry), as I remember from my days at VLSI Technology, is that you always have the wrong amount of capacity. If you don't have enough business, the depreciation on the fab that is not being covered by revenue from wafers kills you. If you have too much business, then your customers get put on allocation and hate you because you are limiting their business. Altering capacity takes a couple of years to build a new fab, meaning that you have to be very aggressive: typically, you have to pull the trigger on a new fab in the depths of a semiconductor downturn, before it is obvious that you will be able to fill it. There’s a famous saying attributed to Robert Palmer, CEO of Digital at the time: “Building semiconductors is like playing Russian roulette. You put a gun to your head, pull the trigger, and find out four years later if you blew your brains out.”

Intel would have to win big customers for their fab, most of whom are competitors. I know from the experience of spinning Compass Design Automation out of VLSI that customers won't ignore that. LSI Logic, VLSI's big competitor in ASIC, really wanted to purchase our datapath compiler, but Wilf Corrigan, the CEO at the time, vetoed it, mostly on the basis that if resources got tight then LSI would lose to VLSI. Intel has also hinted in earnings calls that they may use foundries for more of its business. But most designs aren't worth moving to a different fab. Xilinx, for example, builds older arrays at UMC and probably even Toshiba, an earlier partner.

So, for a 2021 prediction, Intel will not spin out its manufacturing. It will get 10nm working. It will continue to use foundries but it will also continue to manufacture its most important processors in-house.

For the foundries, this is the year of 5nm in high-volume manufacturing (HVM) and 3nm for announcements, test chips, and some risk production. Samsung's 3nm is a gate-all-around (GAA) process. For full TSMC roadmap details, see my post TSMC Technology Symposium: All the Processes, All the Fabs. Here is one quote from that post:

N3 is "the most advanced logic technology in the world". It is still a FinFET process. It will have a 10-15% speed improvement (versus N5), 25-30% power reduction, 1.7X increase in logic density, 1.2X increase in SRAM density, and 1.1X increase in analog density. Risk production is planned for 2021 with volume production in 2022.

The next-generation processes (the ones after 2021) all seem to be some form of GAA nanosheet. I expect we will hear lots about this in 2021 but actual processes will be out in 2023 or beyond. There are probably some technical manufacturing challenges, but the biggest challenge is probably to keep the cost down to a point that the process is attractive compared to the current nodes.

Packaging

The last couple of years has seen a huge increase in "More than Moore" integration, building systems not by integrating everything onto a single die, but using advanced packaging to mix and match chiplets from different processes. In particular, using the most advanced node where it is needed and doing interface logic in a prior generation of process. It is not much of a prediction to say that this trend will continue.

The big question about chiplets is whether a marketplace will develop for them. Today, a company like AMD builds a processor out of several chiplets. But, apart from HBM memory which is well-established, it designs all those chiplets itself. It is certainly possible to envision a world where chiplets are treated like building blocks: a microprocessor chiplet from company A, 112G SerDes interface chiplets from company B, a WiFi interface from company C, and so on. I don't see this happening in 2021. It is unclear who would be the broker, to order chiplets and hold them in inventory, and then fulfill orders. Of course, the chiplets could always be manufactured on demand, but that means that they would only be useful in high-volume products. Lower volume products can't justify a leading-edge SoC design, but they could use leading-edge chiplets if they were available. However, by definition, lower volume products don't have enough volume to justify manufacturing their needs on demand, so there would need to be some sort of broker in the middle. In principle, that could be any of a foundry, an IP company, a fabless semiconductor company, a distributor like Arrow, or even a new company set up specifically for the purpose ("Chiplets R Us").

Einstein

I've already seen several articles pointing out that 2021 is the 100th Anniversary of Einstein receiving the Nobel Prize for physics. But as usual, Breakfast Bytes is ahead of the curve. If you read my post Discovery of the Electron, then you would already know that the answer to the trivia question "In what year did Einstein receive the 1921 Nobel Prize for physics" is actually 1922. The story is at that post but, basically, the Noble committee could not make a decision, and so there was no Nobel prize for physics awarded in 1921. In 1922, they awarded the delayed 1921 prize to Einstein (and the 1922 prize to Niels Bohr). It was also awarded for the photoelectric effect, and explicitly not for anything to do with relativity or gravitation. This wasn't the first time the committee had not awarded the prize and then awarded two the following year. They had done it twice before and would do it five times again (the last time being in 1944).

Since the Nobel ceremony is held in Stockholm each year on December 10, the anniversary of Nobel's death, Einstein received the 1921 Nobel prize for physics on December 10, 1922. I discovered that the ceremony is on December 10 when I worked for Virtutech. I flew into Stockholm to visit our engineering team...on December 10. Going by taxi from the train station from the airport to my hotel was a challenge for the driver since a lot of the town center was closed for security reasons.

Salamis

On the subject of getting dates wrong, 2021 is also the 2,500th anniversary of the battle of Salamis, the most significant battle of the era when the outnumbered Greek city-states decisively defeated the invading forces of the Persian Xerxes. Different sources give different precise dates, between September 23 and 29, 480 BC. But wait, you might say, didn't you read about the 500th anniversary last year? 480+2020 = 2500 after all. But that is not quite the right calculation since that assumes that there was a year 0. But there wasn't, 1 BC was followed by 1 AD. So this September is actually the 2,500th anniversary. Although it is probably actually in October, since I don't expect that the date has been adjusted for the Julian to Gregorian calendar switch, which, as I wrote in my post October Revolution, is how the anniversary of the Russian October Revolution is on November 7. In fact, since Salamis pre-dates even the Julian calendar (45 BC) by over 400 years, I've no idea how the precise date was actually derived.

Prime

Finally, is 2021 prime? It does seem to be if you test it with all the small integers. But it is 43×47.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.