Fabless IDMs Redefine The Leading Edge

Large systems companies are looking more like integrated device manufacturers, designing their own advanced chips, packages, and systems for internal use. But because these are not pure-play chip companies, they are disrupting a 10-year cadence of customization and standardization that has defined the chip industry from its inception, and extending the period of innovation without the associated economies of scale.

What’s changing are the fundamental economics that have driven both Moore’s Law and Makimoto’s Wave. Chips are being designed as part of a larger system, and the economics for internal use are very different for a systems company than for a chip company. Costs can be amortized across thousands of servers, or millions of cars, and designs are so customized that they have little or no value for other companies.

Included in this shift are hyperscalers like Amazon and Meta, as well as automotive OEMs like General Motors and Ford, which increasingly are developing software-defined cars. The common thread is a need for highly-customized device hardware that can be tightly integrated with software, as well as a concerted effort to keep the supply chain insulated from future chip shortages. In addition, there is a growing focus on keeping the details of their proprietary technology a secret.

Rise of the vertical integrators
Put in context, these vertical integrators are in various stages of bringing their design and verification operations in-house, while leveraging economies of scale offered by foundries and OSATs.

This is a new wrinkle in the IDM model, which dates back more than a century when industry titans sought to control their entire supply chain. In the mainframe era, IBM was known for developing and manufacturing its own software and hardware, which it used internally and sold to customers. And in the chip world, it’s a vestige of the days when relatively simple chip architectures made it possible for a single company to design, manufacture, and sell an IC.

Former AMD CEO Jerry Sanders once famously remarked that “real men have fabs.” In this new world order, they don’t. And the IDM giants of yesteryear are trying to find a path forward in a marketplace that demands increasingly customized and heterogenous chips and packages, rendering the old paradigm largely obsolete.

“This whole notion of workload-specific computation is really here to stay,” said Shankar Krishnamoorthy, general manager for Synopsys‘ EDA Group. “Basically, what we are seeing now is that everybody is doing what we call the ‘all of the above’ strategy.’ You can scale, you can put your compute and the parts that are really sensitive on the latest node, but then you can have assembled chiplets from a lot of different nodes. The I/Os can be on 7nm, the memory is on 8nm or 12nm. And essentially you build this system in a single package. The evolution to that is happening extremely quickly.”

The companies embracing the trend now are doing so at a time of great flux and unpredictability for the industry. Moore’s Law increasingly is one element in a much larger solution. And Makimoto’s Wave, needs to be applied or amended on a chiplet-by-chiplet basis.

“Makimoto’s Wave has come to an end for now with the continued requirement for custom design solutions to overcome the limits of semiconductor scaling,” said Rupert Baines, chief marketing officer at Codasip. “The pendulum swinging held true for six decades, which was also a period of time when semiconductor scaling worked predictably. In other words, semiconductor design was delivering more and more performance based on successively smaller silicon geometries supporting growing complexity.”

Fig 1: Examples of standardization and customization cycles in the semiconductor industry. Source: Codasip

The balance between software and hardware is also shifting. “Software is still essential, but software is no longer providing enough possibilities for improvements, and demanding computational performance can often not be achieved with conventional, general-purpose processor cores,” said Baines. “Instead, companies are changing hardware to make software run more efficiently, as well as designing their own processor cores to achieve their design goals.”

Synopsys’ Krishnamoorthy agreed. “One of the most powerful ways to deliver system-level performance is at the intersection of software and hardware,” he said. “The system-level performance is at the workload level? There are many different ways to get it. You can double down on the software and keep the silicon relatively simple. Or you can bring the complexity further down the stack. Co-optimization across the spectrum is really where most people are headed, especially the system companies. The automotive companies and the hyperscalers are looking at this as a full-stack problem, not a silicon problem and a software problem.”

The end result is something novel for the industry. “If Makimoto’s Wave still holds, we should already be in a period of standard solutions but that is not happening,” said Baines. “Instead, we are seeing something new, which is the vertical integration as system companies develop their own in-house SoCs. This started with Apple designing chips for iPhones, but now we see many more companies doing this.”

Specificity, supply, and secrecy
At least part of this is being driven by the fact that good-enough solutions for many of these systems don’t exist, and in most cases there isn’t enough volume to justify the investment in developing them for a broader customer base. Michiel Ligthart, COO at Verific, said that in some cases, a vertical integrator begins exploring the chip world because they have a specific power-related requirement that is not easily served by existing products.

Further, Matt Jones, vice president of strategic marketing at Rambus, said in most instances these companies are focused on the optimization or protection of the application or workload-specific functionality of their devices.

“In most cases, there is a strong correlation for the ‘why’ across a diverse set of applications,” Jones said. “The slowing of Moore’s Law in a time of insatiable growth in data use and requirements continues to drive distribution of compute and storage resources across applications. In these increasingly heterogeneous architectures, we see an increasing number of specialized processing and storage elements becoming more purpose-built for specific applications and workloads. In-housing devices enables companies to optimize solutions for maximum performance and efficiency, while protecting high-value intellectual assets that drive differentiation.”

Sanjay Malhotra, vice president of corporate marketing at SEMI, agreed, noting that automotive OEMs continue to design products that require even more chips while dealing with ongoing chip shortages. “If they do the designs in-house and go to a foundry to have the chip fabricated exclusively for themselves, not only would it have that proprietary label for them, but it would also ensure a captive supply.”

At the same time, these systems companies also rely on proven IP solutions for memory and I/O interfaces to shorten design cycles and meet critical time to market targets. Or put differently, one solution does not fit all applications, but some of the pieces are re-usable across many applications. That, in turn, has caused imbalances in the supply chain as demand spikes, a problem that is particularly evident today in the automotive supply chain.

“They end up with chip shortages,” said Marc Swinnen, director of product marketing at Ansys. “And while they’re scrambling to try and catch up, they’re finding that Tesla, which does control its chips, is edging ahead. Everybody’s worried they’re going to be left in that wake.”

To be sure, vertical integration is not a given for every company or every hardware need within that company, and there are various reasons why a particular organization may choose to pursue a different path. For example, Stefan Poledna, TTTech Auto’s CTO, noted that the economics of manufacturing chips may make it so automotive OEMs focus chip design efforts primarily on hardware with the largest volume of application.

“I am skeptical that it is practical to reduce the volume of the chip by producing a particular ‘Brand X’ chip in, say, only 1 million units because that is the volume of one car model,” Poledna said. “This puts the car manufacturer at a disadvantage compared to a chip from a manufacturer that can be installed in 30 million vehicles, where they have considered a feature that is important to them in the co-design.”

Impact
Is there a chance that the trends toward hyper-customization and secrecy could lead toward a slower pace of innovation if companies are less willing to collaborate with one another?

The answer isn’t clear. Verific’s Ligthart says this is possible, though he does not think the end result will amount to something visible at the macro level.

Friedhelm Pickhard, chief growth officer at TTTech Auto, points to the opposite trend in automotive technology, called horizontalization, which creates a more competitive and innovative environment. “A car manufacturer can use a uniform software platform to integrate different chip modules with different focuses, such as safety, high-performance compute or DNNs, and thus also run applications from different manufacturers,” he said. “If this software platform is both hardware-agnostic and application software-agnostic, a given chip can run software from a specific vendor, but any other software can also run on it. So performance and speed are the same, but the car manufacturer is not dependent on any particular hardware. This stimulates competition and innovation and will lead to more cost-effective solutions in the long run.”

In the shorter term, the most visible impact of vertical integration has been a change in the mix of customers seeking services related to various aspects of chip and system design. Rambus, for example, said it has a more diverse mix of customers as a result of vertical integrators entering the customer base. Similarly, Samer Bahou, director of marketing communications at SEMI, said membership from hyperscalers and automotive companies has increased dramatically in the past few years as these companies get closer to the semiconductor supply chain.

There has been a slow but steady ramping of these changes over the past three years. Ligthart said that more systems companies have expressed interest in working with Verific, though it wasn’t like somebody turned on a switch. “The emphasis on nicheness and customization from vertical integrators is giving rise to related trends in the EDA world. The trend toward in-house semiconductor development is real. We all see it around us, and it is often referred to as ‘bespoke silicon.’ Less visible, but equally interesting, is that this trend goes hand-in-hand with ‘bespoke EDA’ development, which is tailor-made EDA applications to solve domain-specific challenges. Bespoke EDA provides an edge for bespoke silicon.”

Kam Kittrell, vice president of product management and business development in the Digital & Signoff Group at Cadence, said hyperscalers and others have changed the market dynamics around bespoke silicon, which he believes is a renaming of custom ASICs. “They’re making their own silicon or customizing their own silicon because they want to replicate these cards into the machine and put 10,000 machines in a data center, put 10,000 data centers on the planet, and deliver value added-software to their customers. It’s a change in economics that has brought the ASIC back. A few years ago, it was mobile devices driving the market. Next, 5G really changed the game, and AI followed. People are making bespoke silicon, making hardware to drive a specific software stack. The vertical integrators are doing all four things at once, and other successful companies are following suit.”

Vertical integration also has remixed competitive dynamics throughout the ecosystem. Competitors who spent decades fighting for market share now must collaborate in order to survive. Intel, for example, is pursuing a foundry strategy called “IDM 2.0,” which involves expanding its own chip manufacturing capabilities as well as providing manufacturing services to fabless companies. CEO Pat Gelsinger said earlier this month he would be “thrilled” to manufacture chips for competitors like AMD.

“It’s essentially saying, ‘We are your competitor as well as the customer,’” said SEMI’s Malhotra. “We may see more of these cross-connections, depending on the need for a certain process, technology, and capacity.”

Conclusion
How long the reign of the vertical integrators will last remains to be seen, but most experts believe it will not be fleeting. “I believe this is a direction or vector shift as opposed to a short-lived trend,” said Rambus’ Jones. “The drive for increased data computation, usage, and storage continues to outstrip the ability for traditional system architectures to keep pace. More heterogenous models, with workload-specific compute and storage resources will, I believe, continue the drive to take devices in-house as companies seek to optimize and protect their differentiating algorithms and functionality.”

Could these companies even usher in a new phase of standardization in the future? “We will need a new driving force to move out of the time of customization this time,” said Codasip’s Baines. “So far, it is not clear what that would be.”