Transformational Opportunities Coming To Semiconductor Manufacturing

During the GSA US Executive Forum in September 2024, a panel discussion brought together Marco Chisari, EVP from Samsung Semiconductor, Jeff Howell, Global VP for High Tech at SAP, and John Kibarian, CEO of PDF Solutions.

The purpose of the discussion was to compare and contrast the perspectives from one of the largest global semiconductor companies with that of the most widely used enterprise application solution and that of the leading manufacturing analytics provider on three topics:

1. How is the industry responding to the slowing of Moore’s law?
2. What are the operational and business model impacts on the semiconductor supply chain?
3. With governments’ desire for disaggregated supply chains, how do we overcome the resulting challenges?

Demand for advanced computing is strong, fueled by AI, cloud and the electrification of the economy. Yet Moore’s law is slowing down and more computing power will not be achieved by increasing transistor density only. The data in the chart below highlights some interesting shifts and transformational opportunities. With finFET, gate-all-around (GAA), and backside power, innovators continue to find ways to scale.

Fig. 1: Lithography scaling at seven-, five- and three-nanometer continued the conventional 30% shift. Over the next 12 years, it could become more than 30%. (Source: IMEC)

For years, the cost per transistor was decreasing. Then claims were made that it had not only stopped decreasing, but that it was going up. PDF Solutions proprietary analysis shows that the claim may not be supported by the data. The chart below projects that the cost of transistors will not go down and may go up just a bit. This is changing the way industry operates. This is pushing semiconductor designs in new directions, exploring 3D, chiplets and complex hybrid packages.

Fig. 2: An analysis of transistor costs verifies the cost of transistors did not change. (Source: PDF Solutions)

For Chisari, what the industry is doing now from a manufacturing point of view is first and foremost to shift to the next generation of transistor architecture, which is a gate-all-around. “I think it will allow some very interesting new things and will help to maintain the progression of performance. The most complex aspect of this shift is related to the economics of this change. Power and performance will continue to increase, but the cost per transistor effectively has flattened and potentially could actually increase as the cost of the equipment is actually going up significantly.”

What is going to happen from the perspective of cost is that the industry is really going to take a system-level approach. And at the system level, of course, the first solution is chiplet, which means also looking at connectivity. Because, ultimately, what we are going to try to minimize will not just be the cost per transistor, but the cost per bit and the cost of bit movement. And the cost per movement is the cost of the bit and the transistor inside the chip, but also more importantly, the cost of the bit going in and out the memory and the compute unit. And this is really the connectivity. “I think we will see an enormous amount of innovation from that point of view, from a connectivity point of view, at the electrical level, but ultimately at a silicon photonic level.”

Industry organization is also changing. In the early days of the semiconductor industry, everything was defined for final test, yields were high while final tests and packaging were simple.

Today it is a much more complex world. With chiplets comes more test insertions and components and a need to share data up and down the supply chain. Today complexity is at every level of any electronics system. Even the infrared sensor in an iPhone is 13 components in a package and that is a relatively simple system.

Fig. 3: The illustration highlights the increasing complexity of the supply chain for a single die package versus that of today’s 3D hybrid package. (Source: PDF Solutions)

The supply chain challenges are increasing and finding the right combination of applications to plan and control these extended processes is also a challenge.

Engineers struggle with where this is going, notes Howell, who carefully tracks the evolution of this challenge. He looks at the supply chain and compares it to a bicycle chain; it is linear, and it goes left to right. It worked well in the semiconductor industry that traditionally had an inverted bill of materials. But more and more the semiconductor supply chain looks like an hourglass with two distinct supply chain models – design and manufacturing. The traditional applications like 60-year-old material requirement planning (MRP) systems, even with the addition of optimization techniques looking to optimize the flow of materials in the supply chain, remain a black box and are not necessarily well suited to manage this emerging complex supply chain.

That leads software companies to rethink how to approach the hourglass problem.

In his estimation, it’s going to come in three areas. One of them is AI with a new wave of tools designed to take advantage of generative AI to help provide some explanation on what’s going on behind the curtain. The second is business networks connectivity linking together a lot of these multi-enterprises to provide real-time visibility on what’s going on in the extended supply chain. “We are seeing in Germany the emergence of complex business networks with about 200 companies supplying to the auto industry sharing data on a common platform enabled by SAP,” he says.

“And finally, we need to connect the shop floor to the top floor. There is a lot of manufacturing data available tracking where the lots are and when they are coming out and the scrap rates and yields. There is an opportunity for us to marry up some of that data with business data so we can start connecting what’s going on the shop floor all the way to customer commitments and projecting revenue and margins.

“So, I think we are heading into a new wave, a new class of tools that takes the best of the simplicity of MRP, but then also merges it with the sophistication that we got from the AI, large scale data and optimization innovations.”

Manufacturing is something the semiconductor industry wanted to forget about for decades. It’s now front and center and important as the slowing of Moore’s Law changes the way we design systems from an operational and business model, and we need to respond operationally.

At the same time, the semiconductor industry is trying to move manufacturing around the world to new places, as it becomes a strategic government imperative. Every government now believes semiconductors are strategic and they are offering subsidies to put factories around the world. It is creating opportunities and a challenge; the semiconductor industry is good at concentrating capacity in one location and getting to scale. Now the need is efficiency and scale without a concentration of capacity. Leveraging AI is one of the best ways to accelerate learnings and efficiency across the supply chain.

In conclusion, there is an opportunity for us as an industry to go back and say, if we can think about ourselves differently, if we can reinvent the way we look at manufacturing, there’s a tremendous amount of opportunity because the raw numbers suggest we waste the most expensive assets in the world. The most expensive fabs in the world actually produce revenue generating products at best 60% of the time. That’s one of the few industries in the world where we are at that level of productivity out of something that expensive. This creates a huge opportunity for profitable growth across the global semiconductor industry.