Wally Rhines: Predicting Semiconductor Business Trends After Moore's Law

I recently attended a webinar presented by Wally Rhines about his new book, Predicting Semiconductor Business Trends After Moore's Law. Wally was the CEO of Mentor, as you probably know. Now he has the title of CEO Emeritus, something I've never seen before. In fact, I don't recall ever seeing "emeritus" (or "emerita") except in an academic context.

You can download a free copy of his book on the Books page on Semiwiki. (While you are there, you can also download a copy of the book Dan and I wrote, Fabless: The Transformation of the Semiconductor Industry. Or my book EDAgraffiti that Wally described as the "best book ever written on the EDA industry".) I think you probably have to register for Semiwiki to get to this page. If you don't want to do that, all the books are also available on Amazon, but you have to pay.

I always love listening to Wally's presentations since they are heavy in data, often counter to conventional wisdom. For example, everyone says that the semiconductor industry is consolidating, but it is (or at least was in 2016) de-consolidating. See my post Merger Mania on Wally's keynote at that year's GSA Silicon Summit. In the book, he goes further (in chapter 5) and shows how industries often peak with a large number of companies and then drop to exactly three main competitors. Examples he gives are:

• EDA: Cadence...and I hear there are a couple of other big ones (from a peak of three...the top three companies in EDA have about 80% market share, going back to Calma/Applicon/Computervision via the DMV)
• Disk drives: Western Digital, Seagate, Toshiba (from a peak of 85)
• DRAM: Micron, SK Hynix, Samsung (in 1993, the top three had 40% and now they have 95%)

Semiconductor in general has a lot more than three major players today, but the market share of the top 10 (see the chart below) has been remarkably constant:

If you graph the top 50 semiconductor companies, their share has declined over time. So Wally's point that the semiconductor industry is de-consolidating, not consolidating, seems to be correct.

Webinar

Wally started the webinar saying that he's taking a longer view than the next 12 months so the fact that we are in very unusual times since he wrote the book is fine. He has tried to capture trends that stood out over the years that have been reproducible. The book starts with Moore's Law, for example, but Wally pointed out that it started even earlier. Morris Chang at TI (yes, the Morris Chang that went on to create TSMC) found out how rapidly yields improved. That meant that they could predict the price per unit far into the future as a function of the number of units shipped, so they could use forward pricing since they knew what their costs would be. And those lower prices would drive yields down the learning curve even faster. So for Wally, it is all the transistor learning curve, measured by total transistors shipped, that drive both yield curves for individual processes and parts, as well as Moore's Law looking to the longer term. The graph below shows the revenue per transistor from 1954 to 2019:

If you are buying SoCs in advanced nodes you might dispute the above graph. Your costs per transistor are not declining at 32% per year. But this law applies to all semiconductor components, and the cost is made up of a mix of products: not just SoCs, but DRAM and 3D NAND. If you split memory and non-memory it is "evident that cost per transistor and cumulative unit volume for memory are way ahead of non-memory". The reality of the semiconductor industry as a whole is that well over 99% of all transistors are memory.

One big constant of the semiconductor industry is that you buy square inches (or acres) of silicon at a fixed cost. A mixture of increased density, learning curves, bigger wafers, and more, means that the revenue per area for the whole industry is remarkably constant, at about $32/in² as shown in the graph below:

I've heard Wally talk about Gompertz's Law before. This is the official name for what everyone knows as the "S-curve". Benjamin Gompertz described it in 1825. It shows things like the uptake of mobile phones, propagation of a disease, or population growth. It is described by the formula over to the right, but qualitatively you end up with a curve where "only a few people want one, then everyone wants one, and then almost everyone already has one". At the start, there is the highest percentage growth rate (starting from a small base), in the middle the growth rate is linear, and then there is maturity with a lower growth rate.

This has applications in all sorts of areas. Here's one you probably haven't thought of: project management. If you are debugging software or verifying an SoC, in the middle linear part you can predict when you will finish. But Gompertz comes to get you as your efficiency falls off and the project is much later.

The above graph reminds me of an aphorism about software development: the first 90% of the software takes 90% of the time, and the other 10% takes the other 90% of the time.

Profitability

It turns out that there is not much correlation between profitability and size for a semiconductor company. What matters is specialization. In the first half of his career, Wally worked for Texas Instruments (TI) and ended up running the whole semiconductor division. In that era, people used to say that "TI produced everything...except money". From the late 1990s, it shed various businesses and made some acquisitions (notably National), focused on analog and power, and became the most profitable semiconductor company in 2017. It's not just TI, NXP executed a similar strategy (to focus on automotive and security).

Gross margin in the industry is driven by the difficulty of change. People stick with whatever they designed-in for a long time. FPGAs have maintained consistent profitability since once you pick the FPGA architecture it is hard to switch.

EDA

Cadence is in the EDA industry. I think it was Mike Santorini (now at Mentor/Siemens) who came up with the great image that EDA is the waterskier behind the speedboat of the semiconductor industry. If the speedboat is going fast, waterskiing is great. If the boat slows too much, we sink into the water. That's because semiconductor R&D is a constant 14% of semiconductor revenue, and a 1/7th of that goes to EDA. So EDA is basically 2% of semiconductor revenue, rain or shine. If you look at the graph below, you can see it doesn't fluctuate too much, especially recently.

There is a whole chapter in the book on the competitive dynamics of the EDA industry (chapter 7). I'll summarize that in a future post for those of you who don't download the book for yourself.

What Is the New Transistor?

Gompertz's graph for transistors shows us that we will need a new device sometimes, but probably not during most of our careers. As I wrote in my post Exponential Growth, nobody has an intuitive feel for exponentials. Here's a statement Wally made that is both true and completely surprising:

We have hardly produced any transistors compared to what we are going to produce.

Exponentials have this feature that the later part is so enormous that it dwarfs all the earlier part put together. For example, an AMD Epyc Rome has nearly 40B transistors. I'm guessing that that is more transistors than the entire semiconductor industry manufactured in its first few decades. It also leads to some paradoxes where the fastest way to do something may be to wait for the technology to improve.

Table of Contents

Just to whet your appetite to read the book, here is the table of contents:

• Introduction
• Tools for Predicting Semiconductor Trends
  • Chapter 1: Understanding the Learning Curves
  • Chapter 2: Constants of the Semiconductor Industry
  • Chapter 3: Moore’s Law Is Unconstitutional
  • Chapter 4: Gompertz Predicts the Future
• Competitive Dynamics in the Semiconductor Industry
  • Chapter 5: Consolidation of the Semiconductor Industry
  • Chapter 6: Specialization in the Semiconductor Industry
  • Chapter 7: Competitive Dynamics in the EDA Industry
  • Chapter 8: Value Through Differentiation in Semiconductor Businesses
  • Chapter 9: Specialization Inhibits System-Level Optimization
  • Chapter 10: Design Automation for Systems
  • Chapter 11: International Semiconductor Competition
  • Chapter 12: The Future

Watch the Replay

You can watch a reply of Wally's webinar.

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