Intel's Investor Conference: Now Leading with Datacenter

Intel had its investor day recently. I'm not going to talk about Intel as an investment. However, as the largest semiconductor company, with 98% market share in datacenter processors and a leader in process technology, there is plenty to take a look at. Anything in "quotes" means that it is a verbatim quote from either the presenter or the presentation during the morning.

Client Computing and Datacenter

Historically, Intel has used the chips for its highest-volume PC products to drive its process technology. These days, that means what Intel calls all-in-one devices. These basically compress all the functionality of a desktop PC into a very compact form factor. There were two reasons for doing this. The first is that these were the highest-volume products. It is one of the truisms of semiconductor manufacturing that if you want to get a high yield, then you have to run a lot of wafers. The second reason is that the chips were not excessively large (with regard to area). Very large chips yield poorly for defect density reasons, so running very large chips on an immature process is expensive.

So Intel would bring a new process up with notebook and all-in-one chips; then, a couple of years later, as the process matured and capacity became available, they would run server chips. These were much bigger in area and shipped in much lower volume. They had very high margins, so they were an attractive business, but big chips in lower volumes were not attractive as a process driver. Anyway, datacenter chips will the first on 10nm++ (3rd generation 10nm) and on 7nm.

In the terminology of Intel's business units, they would lead with the client computing group (CCG) and follow with datacenter group (DCG).

Diane Bryant, the head of Intel DCG, says that this is changing going forward. They are now "datacenter first." DCG's revenues are still lower than CCG but CCG is stagnant, with revenues maybe eeking out a tiny bit of growth or even shrinking. For sure, the future paradigm for computing is mobile and datacenter. Of course, Intel basically failed to get into mobile and participates only through a few modem wins (including, apparently, some iPhones). However, they dominate datacenter completely. AMD has almost no share in the x86 datacenter market, and 64-bit ARM has, so far, failed to get any significant traction. Intel still has 98% market share in the main processors in datacenter servers.

Another big trend is that enterprise servers are flatlining; that is, companies are building and equipping datacenters by themselves. The main reason for this is that more and more computing is moving to the cloud, which is in a different bucket and is growing fast, 25% last year, and "double digit growth" going forward.

They are trying to expand out from that position to capture more of the other parts of the datacenter, such as networking, silicon photonics. And their 3DXpoint memory, that is now renamed Optane. They also see artificial intelligence (AI)  as an opportunity, both to sell servers and to sell Altera FPGAs (now PSG).

High Growth Areas

AI is one of four areas that Intel calls strategic bets. They are:

• Artificial intelligence and deep learning
• Automatic driving and ADAS
• Virtual reality and merged reality (Intel's name for augmented reality)
• 5G (base stations and modems)

There is no doubt that these are big markets. However, in only a few of them have Intel shown that they are competitive. Intel's one enormous advantage is that it has the x86 instruction set. AMD does too, of course (in fact, technically the 64-bit x86 instruction set is an AMD one, not an Intel one) but AMD doesn't have the financial muscle to compete in more than a few areas. When limited-function computers for internet access first came along, I was at VLSI Technology, and we had a joint program with Intel to put 386 cores along with our ASIC technology to build SoCs for that market. But that market turned out not to exist and the project was canceled early on. One of the big debates was whether the x86 instruction set was a requirement. For a long time, there were people who argued that Intel would dominate in smartphones, because people couldn't run Word and Excel, otherwise. But two things happened. One was that Microsoft ported to ARM. But secondly, it turned out nobody wanted to run Word and Excel on their smartphones.

In the above chart, on the left, the big areas are where Intel is either strong (x86 CPUs), or has a good chance of pull-through because they own the CPU socket (FPGA accelerators from PSG aka Altera, networking). The memory Optane/3Xpoint memory may turn out to be significant, but it is a long time coming. Over to the right, modems are a big market, but Intel has small share. Some parts of IoT (such as ADAS) are clearly going to be big markets even if they aren't big yet; others for the mythical internet-enabled toaster probably are not. The challenge for Intel is that datacenter has strong and robust growth, both as a market and for them, and it is already big. PC CPUs are flat, to probably declining. Other areas are fast-growing, but that doesn't necessarily mean Intel will be growing fast in those markets, plus it starts from a very small base, even if gets good traction.

Process Technology

One of the more contentious areas of any semiconductor manufacturer's presentations is whether Moore's Law is slowing. Intel showed the above graph (HVM stands for high volume manufacturing) and said explicitly that they are continuing to scale logic at 0.5X every 2 years. The other aspect of Moore's Law, namely the cost to keep on whatever line is being shown, is never mentioned. Since Intel's products have such high margin (especially server processors, where there really are no other credible suppliers) they are less sensitive to cost than the average customer of a foundry.

The above chart shows Intel's claim that they are three years ahead of the competition. One area that I don't consider disputable is that Intel really is much more conservative in its process naming than anyone else, and what Intel has been calling 14nm is very close to what everyone else has been calling 10nm. You have to look at the actual widths and spacings of transistors and interconnect to see that. That was the point of The ASML Standard Node. But it seems weird to put "others" having 10nm only this year. Your phone probably contains 10nm chips and you probably didn't buy it this year.

Intel did show some cost data, but not really in a way that is meaningful. They actually said, "although wafer cost is going up, cost per transistor continues to decline," as if it was some anomaly that wafer cost was going up with process generation. The old rule of thumb was one process node was 15% more expensive per wafer than the previous node, and that has been true for decades. However, there are twice as many transistors per unit area, leaving about 35% cost reduction per transistor, despite the increased wafer cost. However, the costs do seem to be going up faster now, with FinFET and multiple patterning both requiring many more mask steps per wafer, hence the discussions about whether Moore's Law is continuing/slowing/over.

Intel was rumored (or, rather, there was a rumor that there was a possibility) to perhaps planning to announce some new type of transistor, perhaps based on something cross-pollinated from the 3DXpoint technology. But nothing unusual like that was mentioned, nor did they license AMD's graphics technology, another rumor that has been strongly...er...rumored.

Intel has a technology day in November (date not yet set), in which more detail on these types of things are sometimes revealed. The investor day was clearly mostly aimed at investors who had an understanding of what markets Intel sold into but had a limited understanding of the underlying semiconductor technology and economics. I've done my best to unpick what was said for you.

If you want to know more, then there are videos of all the presentations, and also copies of the slide decks, on the Intel website here. But, to be honest, I watched all four hours live so you don't have to watch it all yourself.