A History of Semiconductor IP

I like to claim that I was in the IP Business before the name IP was used for semiconductor components. When VLSI Technology spun out Compass Design Automation as a separate company, in addition to all the EDA software, Compass also inherited what we then called the "library business". We created standard cell libraries (and a few gate-array libraries) not just for VLSI themselves, but also for several other companies, mostly in Asia and Europe. We also created SRAM compilers. Soon after that, Artisan Components was created, so we had the validation of competition that this was a real market. Artisan would eventually be acquired by Arm and transmogrify into what is today Arm's Physical IP business.

Foundation IP

One thing that we discovered was that design rules were treated as incredibly confidential crown-jewel secrets. For example, VLSI printed the design rule document with a broad maroon stripe across each page so that it wouldn't photocopy. Sometimes, the gating item to delivering a library was getting the customer's legal department to sign off on us getting the design rules so that we could start the project. Back then, processes were less varied than today, since the scaling was largely limited by the tolerances of the tools from the equipment vendors. But as a result, all these crown-jewel design rule documents had almost identical design rules. That gave us the idea to come up with Passport, a standard set of design rules which we could use to create libraries that could be manufactured by anyone.

At that point in the evolution of the library/IP business, customers were usually facing a make-versus-buy decision. Designing standard cells does not require a lot of specialized knowledge. Designing SRAMs was fairly straightforward in those processes, too, but most semiconductor companies still lacked the software expertise to create an SRAM compiler that could create SRAMs of whatever size you wanted. They would typically just build each SRAM by hand. The big advantage we had, besides having the software expertise to build compilers, was that we had a team that did nothing but design standard cell libraries. Our customers always seemed to run into trouble since their standard cell libraries were always being created by novices. Creating standard cells was normally an entry-level job in a semiconductor company, and once engineers had done that they wanted to move onto more differentiated silicon. No company won business because their standard cells were better than the competition's.

But they could lose business if they didn't get their libraries out in time, or they had problems. Gradually, the idea that you shouldn't design all the libraries that you use started to catch on. Semiconductor companies redirected the engineering resources to things that mattered to their customers, and great 3-input NAND gates were not on that list.

Processor IP

We were not in the microprocessor business, except as one of the original investors in Arm when it was spun out of Acorn. Initially, it was hard for VLSI to get customers to use Arm processors in their designs (so that we could build the silicon since initially we were the only licensee). The notion that anyone would license a microprocessor from someone other than an actual semiconductor company that manufactured microprocessors already was too alien. So the semiconductor world was divided into semiconductor companies that had their own microprocessors (Hitachi, Motorola, NEC, Infineon, NXP, and so on) and everyone else. As chips got larger, ASIC glue logic gave way to true systems-on-chip (SoCs) and these always involved a microprocessor. The companies in the "everyone else" category all licensed Arm as a way to get one. Over a few years, the dynamic changed. It was easy to get customers to use Arm in designs (but VLSI now had a dozen competitors, too). The "own microprocessor" companies started to discover that their microprocessor was a liability, not a differentiator. Over time, pretty much all of them would license Arm, especially if they had anything to do with mobile where the ARM7TDMI became the de facto standard. Licensing Arm was not a make-versus-buy decision though. Even though it wasn't all that hard to design a microprocessor of that era, everyone realized that they were actually more in the software business. It was compilers, debuggers, operating systems, device drivers, and the like that were the challenge. Arm's ecosystem, not its processors, were really its moat, to use the VC term for making a business defensible.

Interface IP

A big transition was the switch from parallel low-speed interfaces to very fast high-speed serial interfaces. This started first with DRAM with DDR standards. These were challenging to design and even more challenging to verify. Denali started just making verification models, what today we call VIP (for Verification IP). Then they started to build IP to go on the chips. For many customers, this was not a make-versus-buy decision because it was too difficult and expensive to make. Also, from a business point of view, there was little upside in developing standards-based IP internally. Within reason, the only thing that matters about standards-based IP is whether it meets the standard. There is no upside in a DDR interface running faster than the DRAM chips, it's not like a microprocessor where faster clock speed or lower power for the same clock speed is very important.

Over time, other standards-based IP, such as PCIe, Ethernet, and USB were added to the portfolio.

As had happened with standard cells a decade earlier, it became clear to semiconductor companies that they were just wasting resources on undifferentiated silicon if they designed their own standards-based IP. Serial interfaces, and the SerDes interfaces behind them, have gotten faster and faster. The current state-of-the-art is 112G, 112 gigabits per second. This is not something straightforward to design and almost all design groups license SerDes IP from an IP company like Cadence.

System IP

Another area was specialized processors. No company, except in the x86 market, considered designing its own general-purpose processor. That was over and Arm won, although RISC-V may make some inroads. But as Moore's Law slowed, the only way to deliver increased compute was with a processor focused on the application. General-purpose processors could do anything, but didn't do anything in particular very well compared to a processor tailored to the job.

Tensilica processors are based on the Xtensa platform that can tailor the processor in many ways. But many system companies tend to be working on something higher-level, like audio, or vision, or radar. They do not want to worry about tailoring processors and it's not their area of expertise. They are experts on audio, or image recognition, or radar imaging. As a result, Tensilica processors are usually sold as more of a system solution, including not just a processor already tailored to the application but also application software too. For example, the Tensilica HiFi series has over 120 licensees, at least partially because every codec you might want to use already runs on it, along with other specialized audio solutions like Dolby Atmos. These are true system-level IP solutions.

Going back to VLSI Technology days, one company we worked with was called Symbionics. They had specialized blocks and software for implementing GSM (now often called 2G, although it wasn't then). So this was another example of true system-level IP. As it turned out, when Cadence created a design services subsidiary called Tality, they acquired Symbionics. Unfortunately, Tality was a victim of the 2001 technology downturn, when all their contracts got canceled. When Tality was riding high before the downturn, it had turned down a nine-figure offer to sell Symbionics. After the downturn, the whole business fell apart, and the employees of Symionics ended up being laid off (along with many other employees of Tality).

I expect to see a lot more system-level IP in the future. A lot of IP is only useful when it is bundled with the interfaces, radios, and software required to make it work. These all need to be designed to work cleanly together so that implementing a system in that domain, be it automotive, data center, a 5G basestation, or something else, then all the building blocks are available and have been designed holistically.

Learn More

See the Cadence IP product page.

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