Building Smarter, Better Fabs

Battling labor shortages, faster ramp rates, and data overload, the process of designing and building greenfield fabs requires a combination of tech tools, failing earlier approaches and superior planning from day one.

The complexity and scale of semiconductor fabs is skyrocketing as is the capital cost. Chipmakers are looking to ramp multibillion dollar fabs faster despite the hurdles of labor shortages, material delays, and intense competition.

One group of experts at SEMICON West (see photo below) described ways that augmented reality, AI, synchronized data access and digital twins can help make smarter fab design and construction a reality.

Fig. 1: [L-R] NVIDIA’s Chen; Exyte’s To; Bechtel’s Smith; Jacobs’s Fullam; and Deloitte’s Mitra. Source: Laura Peters/Semiconductor Engineering

Part of that reality is recognizing design and construction modifications happen simultaneously. “It is like flying a plane while you’re building it,” said Evann Smith, digital solutions manager at Bechtel. “When we think about construction of capital projects, we think of a locomotive. It’s linear, it’s on the track moving full steam ahead, and we know what our destination is. That is not the way of the semiconductor fab. We’re talking about dynamic and fluid projects that present both a challenge and an incredible opportunity.”

That opportunity became clear to fabless semiconductor firm NVIDIA two years ago when it began developing a digital twin platform for fab infrastructure applications, which Samsung Electronics expects to prototype in 2025. “Samsung in Korea is starting to build this digital twin for its (fab) infrastructure, which was actually presented in public at our GPU technology conference this March,” said Jerry Chen, director of business development for Manufacturing & Industrials at NVIDIA. “Part of our mission is we’re going to make AIs that are able to understand physical spaces and cyber-physical systems. Now, our purpose is to build platforms for other people to build end-user solutions off of them.” (See Seokjin Youn of Samsung Electronics’ video here).

To better address chipmakers’ priorities and needs, Ricky To, senior manager of Data & Digital Delivery at Exyte, called for a change in mindset around fab construction. “Compared to the software industry, which is very iterative, the fail-fast approach is really taking off,” he said. “But in the construction or design engineering industry, we typically iterate when a project finishes.”

By changing the approach from failing slow to failing fast, fab builders can shift the bulk of the risk of changes from late to early stages while reducing material waste and cost of delays. Better use of historic and current data is central to this shift. “We have a real-time data synchronization and our cloud architecture facilitates that. There are opportunities to reduce interfaces. So a lot of times when we build these projects, we throw things over the fence at each other,” said Bechtel’s Smith. “But because the opportunity for influence is highest early, from a data perspective we need to be integrated from the beginning with transparency amongst partners, to appropriately impact that curve.”

What goes into a fab?
According to Deloitte, a leading-edge fab capable of producing 20,000 to 25,000 wafers per month costs between $10 billion and $12 billion, with a building cost of $4 billion to $5 billion. It takes two to three years to go from groundbreaking to first silicon production, but a relatively small portion of the manufacturing site is the fab cleanroom facility itself. Deloitte also estimates that building such a fab requires 6,000 to 8,000 workers.

The tight labor market also means semiconductor competes with multiple industries for a skilled workforce, especially when other areas of the economy boom at the same time. “Demand is going up,” said Exyte’s To. “Everything is in demand across communication systems, the automotive industry, computing, and storage. That means that we need more facilities, more resources. And this will, of course, incur more cost and energy consumption, as well. That affects construction because, basically, we’re going to need more people with the right skill sets, we’re going to need more equipment, and software.”

In 2023, investments poured into semiconductor facilities including fabs, assembly and testing houses. SEMI estimates more than 60 new fab projects will begin construction through 2030, but capital expenditures typically are closely aligned with global economic conditions, which explains tool delivery delays after groundbreaking is announced. Also, because the facilitization (tool installation) part carries the bulk of the cost of a fab (50%), shells can remain dormant during semiconductor down cycles. SEMI further estimates that $36 billion will be invested in construction projects in 2024.

From a design perspective, a simple way to understand fab construction is that they generally work from the bottom to the top, and from the outside to the inside. “Typically, the early release packages will focus on piles/foundations and undergrounds, then concrete and shell, including building steel framing, then the interiors, and finally the MEP scope, the infrastructure and utilities that serve the toolset,” said Paul Fullam, director of technology, EMEA, Advanced Facilities at Jacobs Engineering.

Implementing tech tools
Ideally, smart technologies such as augmented reality (AR) tools (see figure 2) and autonomous robots are employed in large construction sites but enormous models, bandwidth limitations, and maintaining uninterrupted network connectivity can prove challenging.

Fig. 2: Exyte’s construction engineer uses augmented reality to overlay design components and actual piping on site. Source: Laura Peters/Semiconductor Engineering

“Augmented reality is deployed on all of our sites. You hold it up, look at your construction site and it shows you what’s there, and then it overlays the model of what’s to be installed. So our superintendents in the field can plan their work and they love it, “ said Bechtel’s Smith.

Augmented reality is already used on-site, and it’s a brilliant tool,” said To. “But there are limitations. We constantly need to work offline because our facility models are massive, meaning we need to slice up the model, then load them into an iPad. Unfortunately, the limitation is the hardware itself. So if we need to move into another space, we remove the old model and then load in the new model. This becomes a very long, painful process if there is slow or interrupted internet connectivity.”

He added that robotic scanning, using a Boston Dynamics robot, for instance, can scan while avoiding obstacles once it’s trained and learns the environment. “Robotic scanning is a great technology that’s been used for many years, but it’s extremely, extremely slow. So we’re looking at automating an entire process so we can bring information faster to our design model. Then this scan data is streamed into a platform where it can remove the noise and auto-register, auto-join the various scans that have put in place. Finally, it moves into a platform where it overlays the scans onto the 3D model for comparison to see deviations,” said To.

Fast-tracking fabs
The speed at which chipmakers wish to ramp fabs is not a new dynamic, but complexity and scope have grown with the multitude of sensors, connections, data streams, and stakeholders involved.

“Semiconductor fab construction projects are traditionally fast-track in nature, and subject to overlapping construction and design activities,” said Jacobs Engineering’s Fullam. “Successful execution in this environment requires tight design and construction integration.” He points to two levers chipmakers using to compress project timelines further: leveraging reference designs, that can be used from fab to fab, and off-site manufacturing, building sub-assemblies or systems off-site and then delivering them to the site for integration.

Fig. 3: Reference designs and greater use of off-site (OSM) manufacturing are two strategies aimed at compressing the design/construction timeline. Source: Jacobs Engineering

Trade contractors in the semiconductor space have been using off-site manufacturing (OSM) for a decade or more because it’s in their own best interest. They can assemble substantial portions of their work product in a controlled factory environment away from the site.

“The benefits include increased quality, safety, and reduced trade-stacking constraints at the project site,” said Fullam. “But the dynamic that’s different now is that the level of OSM that our customers are pursuing is not 20% to 30% of the overall job in terms of labor hours, but pushing to 50% and beyond. This drives a new dynamic in design, and how design and construction are orchestrated. Designing for OSM requires that you include specialists who understand shipping splits, understand naval loads if modules are being shipped from overseas, and rigging, sequencing, etc. as part of the design process. These are things you would not typically have dealt with during the design phase.”

Results one can see
On semiconductor sites, there is a disconnected digital landscape in the design and construction toolset, which necessarily includes solutions from multiple vendors including Hexagon, Assemble, Oracle, SQL Server, and many others. “Bechtel uses multiple solution throughout the engineering, procurement and construction process, but the critical thing is they are all connected in a data mesh architecture, providing real-time, secure data synchronization — not data exchange and not end-to-end,” said Smith. “This is synchronized across our systems so that we understand what’s happening, where, when, and how.”

To avoid confusion, shared visibility is at the core of smart design and construction portals. “One key takeaway is to make data visible. There are myriad software tools available that allow live work-in-progress data to be made available broadly, and we have seen benefits to projects where such ‘real-time’ data sharing is leveraged,” said Jacobs’ Fullam. “With respect to BIM, many of our customers and stakeholders, especially operations personnel, process concepts visually. So oftentimes, if you develop a concept model spatially — even to a relatively basic level, to explore multiple different configuration options — it can allow the stakeholders to make key decisions and option selections earlier in the design cycle. That reduces the potential for later churn,” he added. “We have found the ‘rapid-prototyping’ practice that BIM platforms afford to be really, powerful.”

Other panelists agreed. Exyte’s To emphasized that synchronization between the physical twin models as they are being built and large design models can be an extremely slow process. Synchronization must deal with multiple data formats coming from video captures, image capture, laser scanning, etc.

Conclusion
Semiconductor fabs are possibly the most complex and expensive production facilities on the planet. Fabs take two to three years to go from groundbreaking to operation, known as first silicon. By taking advantage of some of the intelligent technologies made possible by the chips themselves, the construction industry is working to compress schedules beyond historic norms. There are growing trends toward synchronous data sharing, AR, and autonomous robot implementation. Alongside of those trends, large-scale hardware/software platforms to process massive data streams may become essential.

Related Reading
Money Pours Into New Fabs And Facilities
Investments boom as countries and companies vie for supply chain security and technology leadership.
Digital Twins Target IC Tool And Fab Efficiency
Virtual representations will improve performance and productivity across the entire design through manufacturing flow, but deployments will vary in effectiveness and timelines.