Last year Dan Nenni and I wrote a book on the semiconductor industry through the ages called Fabless . Actually I did most of the writing and I think Dan thought he had it easy. He just had to get contributed sections from all the companies in the industry. Everyone was eager to do it but nobody was eager to make it a high priority, so getting those chapters was harder work than my getting my part done. My sort-of-predecessor here at Cadence, Richard Goering, wrote the Cadence section. The subtitle of the book was The Transformation of the Semiconductor Industry . The link above will take you to Amazon but unless you really really want to pay for a printed copy (go for it, I get two bucks or something), you can get a free PDF version on Semiwiki . You can also get a free PDF version of my earlier book, EDAgraffiti at the same link (the book that Wally Rhines told me was the best book on EDA, except I pointed out that it is probably the only book on EDA). If there is one theme that runs through both books, it is that economics drives the semiconductor industry. Moore's Law is a law about economics as much as a law about technology. As the economics change, the structure of the industry changes to reflect that. We are in the midst of another change, I believe. The way system design is done changes roughly every ten years. Since the main driver of these transitions is changes in the economics of semiconductor manufacture, that is five process nodes or so. The boundary between what is done in the system companies and what is done in the semiconductor companies moves. In the 1970s, all the knowledge about how to design semiconductors was in the semiconductor companies. This was the era of the "tall thin man", people who understood everything from process technology to layout and simulation. But process technology moved on. The publication of Mead and Conway's seminal book Introduction to VLSI Design created a cohort of designers who didn't understand process technology in depth, and processes go so complex that they had their own specialists in what was known as TD, technology development, who knew nothing about design. Computer scientists suddenly learning how to design chips rapidly produced first primitive EDA tools and eventually an entire industry. The dominant methodology for system design was ASIC (as an aside, at VLSI Technology we tried to call it CSIC since designs are actually specific to customers not applications but ASIC was a lot easier to pronounce). The part of the design close to the system, front-end design, was done in the system company. The part of the design closest to the process, physical design, was done at the semiconductor company. Netlist went one way and back-annotation went the other. The next step was that the system companies resented being dependent on the semiconductor companies and wanted to get more control of their destiny. They wanted to go all the way to layout themselves. This was partially a result of the EDA industry maturing enough that system companies could get the tools, and also that the IP business started to come into existence supplying basic silicon structures such as standard cells and memories. But especially the beginnings of standalone foundries. This style of business was known as COT, or customer-owned tooling. The word tooling is semiconductor jargon for the masks and, in ASIC, the owner of those was the semiconductor company, not the customer. With the beginnings of foundries and the fabless ecosystem, the focus of the book I mentioned at the beginning, the customer had complete control. But semiconductor design became more and more difficult and gradually system companies, especially the cell-phone companies experiencing explosive growth, let design move back to the semiconductor companies. For example, Nokia even sold a semiconductor design team to what is now ST Microelectronics. This worked well for a while but, in the last few years, system companies realized that they couldn’t have differentiated products if they just wrote software to run on the same semiconductor products also available to their competitors. The entire system needed to be designed concurrently, all the way from application software down to the transistors. This application-driven development process Cadence calls System Design Enablement (SDE). This is a big change. No longer do system companies just do software, and semiconductor companies just do semiconductors. Change are taking place in both directions, with semiconductor companies creating entire software stacks. One reason is that it is too difficult to verify the silicon without the software. GPU companies cannot sell chips without the graphics software running right. Nobody in their right minds is going to tape out a chip for an Android phone without first booting the operating system and probably more. But booting Android is measured in the billions of simulation vectors. This is one of the reasons that emulation is a fast-growing part of the EDA market. Emulators are expensive. No marketing slide from Cadence (or its competitors) ever says that directly. But they are also the cheapest verification vector you can buy—it's a sort of quantity discount—but you had better need trillions of vectors (or what is next, quadrillions?) and increasingly you do. I like to say that the Palladium Z1 enterprise emulation platform is an enterprise server farm in a rack, only faster. It's not totally accurate, tiny simulations that run for only a minute are not worth the hassle, but it's close. But even so, system companies do not want to depend on getting everything from a semiconductor company that is also available to their competitors. As a result, system companies have been moving deep into semiconductor design, hiring semiconductor teams so that the entire system can be optimized together to produce a stack of software/silicon not available to their competitors. It is notable, I think, that all the market leaders in the smartphone industry design their own application processors in-house. Me-too is no longer good enough. So system companies are becoming more like semiconductor companies, and vice versa. Everyone needs to design from application software down to the silicon. Irrelevant P.S.: Rereading this I just noticed that the spellchecker in our CMS doesn't recognize the word "enablement" and proposes to make it "disablement" which is not quite the same thing at all! At least it didn't change it automatically. Another pick is "ennoblement". Arise Sir Chip.![]()
![]()
↧