How GlobalFoundries Plans to Profit from AI Without Making GPUs

From the light and electricity of data centers to cars and humanoid robots, this foundry bets on everything beyond computing power.

In the past two years, when talking about AI, one cannot bypass GPUs.

On September 22, during the GlobalFoundries Technology Summit (GTS), Shankaran Janardhanan, Senior Vice President of GlobalFoundries' Photonics and RF Business, gave another judgment: "The bottleneck limiting AI development is no longer computing power, but connectivity and power supply."

GlobalFoundries has its stance in saying this. As early as 2018, this wafer foundry paused 7-nanometer R&D and withdrew from the competition for advanced processes. It does not make GPUs, nor does it compete for 3-nanometer or 2-nanometer, but instead bets on everything beyond computing power: who will do the sensing, control, and low-power computing needed when AI enters the real world, how data is transmitted out, and how power is delivered in.

That day, five senior executives from the China region, Asia-Pacific, and global business took turns on stage. Topics ranged from cars all the way to data centers and humanoid robots. The scene span is not small, but upon closer inspection, GlobalFoundries is always solving two things: connectivity and power supply.

Let's start with cars.

01.

Cars: Car numbers haven't increased much, but chip numbers have

"For any technical roundtable or summit, it is incomplete not to talk about cars in China." Faisal Saleem, Global Senior Vice President of GlobalFoundries, said right from the start.

Faisal Saleem, Global Senior Vice President of GlobalFoundries

He summarized four changes happening in the automotive industry. First, cars are getting "smarter", sensing the surrounding environment stronger, and there are more connections between car parts, cars and people, and cars and clouds. Second, software-defined cars, functions can be opened, closed, and upgraded remotely, hundreds of MCUs that were scattered across the whole car in the past are being replaced by regional architectures. Third, electrification, over half of new cars in China are some form of electric vehicles, behind which is a fundamental change in the whole vehicle power architecture. Fourth, supply chain resilience, after experiencing chip shortages, automakers need both local supply chains and global supply chains as a backup.

Faisal also mentioned a set of interesting numbers. Over the past five years, the number of cars using GlobalFoundries chips has hardly changed; what changed is that there are more and more chips on each car, and they continue to increase.

Where do the extra chips come from? Cameras, LiDAR, millimeter-wave radar, ultrasound, and cockpit sensors continuously generate data, these data need to be processed and transmitted. Thus, domain sensors, regional sensors, network chips, Ethernet chips, SerDes, and power management chips are all increasing. In the end, it is still "connection" and "power".

After there are more chips, new problems start to appear, which is "no single process can handle everything".

Millimeter-wave radar transceivers and battery monitoring chips have completely different process requirements. In Faisal's view, dedicated process platforms will become increasingly important: FDX (FD-SOI) for low power and RF, BCD for battery management, Gallium Nitride for power.

Additionally, in terms of competitiveness, the core selling point of automotive chips is trust

"Customers buying cars are buying not just technology, but more of a kind of trust." Faisal said. A car needs to drive for 10 years, 15 years, or even 20 years, often working under extreme temperatures, so quality and reliability are the first priority. GlobalFoundries built the AutoPro platform for this, integrating automotive requirements into every link of design, process, PDK, IP, and manufacturing operations.

GlobalFoundries has done automotive for over 15 years, processes cover 180 nanometers to FinFET. In Faisal's words, "Without exaggeration, I can say there is not a single automaker whose cars do not carry our products".

SVP of Asia Pacific Manufacturing and Singapore Site and General Manager Chen Yaoguang

Trust ultimately needs to be redeemed by factories. Over 40% of the annual production capacity of GlobalFoundries' Singapore site is automotive-grade products. Chen Yaoguang, SVP of Asia Pacific Manufacturing and Singapore Site and General Manager, shared the factory's performance in manufacturing quality, production efficiency, and delivery reliability, demonstrating GlobalFoundries' emphasis on automotive chip quality and stable supply.

When talking about AI in cars, GlobalFoundries China President Hu Weiduo believes the core inside cars is still MCU, this is also the direction GlobalFoundries is heavily investing in domestically. Computing power needs to be improved, power efficiency needs to be improved, but "the foundation is still stability, reliability, and security as the key pillars for all automotive-related products".

GlobalFoundries China President Hu Weiduo

02.

Cloud: Same challenges, magnified by several orders of magnitude

From cars to data centers, what GlobalFoundries faces is actually the same question: power is getting larger, electricity and data are getting harder to deliver. It's just that at AI data centers, the order of magnitude has multiplied several times over.

In terms of power, the trend is from 48V to 800V.

Shankaran summarized the changes in data center power supply into three points: higher current density, voltage rises to 800V high voltage DC, higher power density. This is the same principle as cars going from 12V to 48V. When power goes up, one must rely on increasing voltage to lower current and loss.

The trouble lies in that these things restrict each other. Data needs to be transmitted further and faster, power supply must be closer to the rack. GlobalFoundries' trump card here is BCD, Gallium Nitride and other power and power platforms, which are the processes already used in cars.

In terms of communication connectivity, the trend is copper retreats optics advance.

The pressure on connectivity is greater. Data centers are moving from single rack to multi-rack interconnection, GPUs to be connected go from dozens to hundreds or thousands. Single channel rate has gone from 100G to 200G, next is 400G; power consumption needs to be compressed to below 5 pJ per bit.

Shankaran divides interconnection into three layers. Vertical scaling (scale-up) within racks or clusters, copper cables have fallen behind bandwidth needs, currently "copper retreats optics advance"; horizontal scaling (scale-out) between racks and clusters, and cross-data-center, cross-campus cross-domain scaling (scale-across), was originally the realm of light. In terms of form, the current mainstream is pluggable optical modules and NPO (Near-Package Optics), next step is CPO (Co-Packaged Optics).

This is the area GlobalFoundries is most confident in. Its silicon photonics and silicon germanium (SiGe) technologies are both mass-produced, among which 200G/λ has been shipped, 400G/λ and above are in R&D, the roadmap also includes 12-inch silicon photonics. Previously, GlobalFoundries also acquired Singapore silicon photonics foundry AMF. According to Chen Yaoguang, if only pure silicon photonics revenue is counted, GlobalFoundries is already number one globally.

As light and electricity get closer, the final competition is heterogeneous integration, which happens to be where GlobalFoundries is still catching up. Chen Yaoguang said very frankly: "Advanced packaging is a field where we still need to further expand capabilities. We have already completed half, there is more work to do".

In the parts already achieved, TSV (Through-Silicon Via) has been in mass production since 2014, SLATE wafer-to-wafer (W2W) hybrid bonding also has mass production capability. In the future, GlobalFoundries also plans to further promote chip-to-wafer (D2W) hybrid bonding technology, and explore integration with Gallium Nitride and silicon photonics technologies.

03.

Humans: Car foundation, plus AI brain

"Data centers will have further growth, and artificial intelligence will step out of data centers, enter the real world to promote physical AI development." Those who said this are MIPS CEO Sameer Wasson, MIPS is now a company under GlobalFoundries.

Sameer sees physical AI as a closed loop: Perception, Thinking, Action, Communication. GlobalFoundries' next investment will revolve around these four links.

Humanoid robots are the most complete sample of this closed loop. GlobalFoundries gave two numbers at the summit : The semiconductor usage of a humanoid robot is 3 to 4 times that of a car; to achieve dexterity operations close to humans, more than 40 degrees of freedom are needed.

However, if you break down the robot usage, it is still the car set.

Perception relies on vision, acoustics, LiDAR, millimeter-wave radar, and motion position sensors; Thinking relies on multi-modal AI and sensor fusion, requiring low latency; Action relies on efficient, precise actuators and motor control; Energy relies on battery management and power management chips.

This list is almost no difference from a smart electric vehicle's chip list. Even the architecture thinking is the same, GlobalFoundries labeled for humanoid robots "Distributed Intelligence and Regionalized Architecture", and cars are walking the same road.

Chen Yaoguang also pointed out this point when answering automotive-related questions: "The MCUs we focus on are used not only for electric vehicles, but also for humanoid robots." High voltage BCD processes are the same. He also specially mentioned, humanoid robots need to deal with humans daily, safety is very important. And safety is exactly what the automotive system has polished for over ten years.

GlobalFoundries doesn't make GPUs, but makes "small brains".

The biggest difference between physical AI and data centers lies in the "Thinking" link. Sameer believes, large GPUs, NPUs will certainly still exist, but future systems will move more towards distributed intelligence, the physical world needs large amounts of distributed, extremely low-power intelligence.

He gave an example: "Motor control is not a fresh thing, but if it is driven by generative AI or AI, this is a new thing."

This also explains why GlobalFoundries wanted to acquire MIPS. After buying MIPS last year, in the past 12 months GlobalFoundries also acquired ARC processor IP, Sameer described "our capabilities doubled".

Based on these IPs, GlobalFoundries launched three computing reference platforms: Inference-native Acies, adopting 100% RISC-V architecture; Automotive and Robot real-time applications Actus, based on FDX process and Arm; and Security-native CPU Aegis, also 100% RISC-V.

Behind this is the expansion of customer boundaries. Sameer stated, chip design companies are still GlobalFoundries' important customers, but the company is also expanding its customer group to Tier 1 suppliers, extending the value chain backward.

A pure wafer foundry starting to provide CPU IP and reference platforms is not common in the industry. GlobalFoundries bets on: in the physical AI era, whoever can integrate computing, sensing, and power supply at the lowest cost will get more designs.

Epilogue: Placing a Move in Guangzhou

Cars, Cloud, Humans three scenarios span a large range, GlobalFoundries puts out is the same set of assets: FDX, BCD, Gallium Nitride, RF, and silicon photonics these differentiated processes, heterogeneous integration capability that seals them together, MIPS and ARC brought proprietary IP, and factories distributed in multiple places.

The last item is especially urgent for Chinese customers. Chen Yaoguang gave an example, 40 nanometer, 55 nanometer processes can be produced at factories in Singapore and Dresden, Germany, customers changing factories do not need to design and verify from scratch. Now, on this map there is one more point: Guangzhou.

Early in September, GlobalFoundries established an office in Guangzhou, and cooperated with Guangzhou Zengxin Technology. From the information disclosed at the summit, Zengxin is GlobalFoundries' manufacturing partner, providing customers with "China for China" capacity choice; technology transfer is led by GlobalFoundries Singapore, initially focusing on the automotive field.

When answering questions, Faisal said, automotive electronics is one of the fields with relatively clear localization demands, customers all hope to "have both a global supply plan and a local supply plan". This article starts with cars, GlobalFoundries' move in China also starts with cars.

Hu Weiduo said a sentence at the start of the communication meeting: "You can see in semiconductor nodes that behind the numbers there are differentiated functions and performance." Outside the GPU spotlight, chips needed by AI are far more than computing power chips, what GlobalFoundries wants to do is exactly these.