GlobalFoundries 发布 FDX Fusion 技术瞄准物理 AI

GlobalFoundries today (Oct. 9) unveiled FDX Fusion in Dresden, a new generation of its fully depleted silicon-on-insulator (FD-SOI) technology aimed at physical AI. The bet is that the next big AI semiconductor market will look very different from the GPU race in data centers.

FDX Fusion targets machines that must sense, decide, communicate, and act—not just compute. The first generation is expected to offer 7-nm-class digital performance and more than twice the density of GlobalFoundries’ first-generation FDX technology. It will also combine RF, analog, mixed-signal, embedded nonvolatile memory, in-memory computing, and dense logic capabilities.

Horstmann said GlobalFoundries expects to invest roughly $1 billion over the coming years in FDX Fusion development, including about $250 million for the initial capability toolset. Demonstrator silicon is targeted for early 2027, with an initial process design kit planned for mid-2027.

The announcement came as GlobalFoundries celebrated the topping-off of SPRINT, its Dresden fab expansion supported under the European Chips Act. The project is adding cleanroom, laboratory, and manufacturing capacity at a site GlobalFoundries plans to use for differentiated semiconductor technologies.

For Manfred Horstmann, senior VP and general manager of European fabs at GlobalFoundries, the FDX Fusion architecture reflects a fundamental difference between the chips needed for data center AI and those required when intelligence moves into cars, robots, drones, industrial equipment, and other autonomous systems.

“Data center AI is extremely power hungry,” Horstmann told EE Times in an exclusive interview before the announcement. “Physical AI is a different class of problem—and it needs more than compute. It needs sensors, actuators, communications, low power, and cost efficiency.”

That’s the idea behind what GlobalFoundries calls STAC—sense, think, act, communicate. Instead of optimizing primarily for maximum digital compute density, FDX Fusion is intended to combine substantial processing capability with the functions needed for a machine to interact with its environment.

From FD-SOI to FDX Fusion

FDX is GlobalFoundries’ implementation of FD-SOI. Its transistor channel sits in an approximately 70-angstrom, or 7-nm, silicon film above a buried oxide. Because the layer is so thin, the channel is inherently fully depleted. “You don’t have to waste energy depleting the channel,” Horstmann said.

That gives FDX some of the electrostatic advantages associated with FinFETs without requiring GlobalFoundries to construct a fin. “A lot of the functionality comes from the substrate itself,” Horstmann said. “So we don’t need to build a fin to get a fully depleted transistor.”

GlobalFoundries is developing the strained-silicon substrate with CEA-Leti and Soitec through a European project that Horstmann identified as “Fast and FAMES.” The partners have built a pilot line in France for the substrate, which uses a strained 70-angstrom silicon film intended to increase electron mobility by about 25%. Horstmann said the collaboration builds on the same consortium that worked with GlobalFoundries on 22FDX.

The new substrate is part of GlobalFoundries’ effort to reach 7-nm-class performance with a simpler process flow than advanced FinFET technologies. “For 7-nm-class performance, we don’t need EUV,” Horstmann said.

GlobalFoundries said it expects the initial FDX Fusion process to require about 50 masks, compared with the 70 to 80 masks Horstmann cited for advanced 2-nm and 3-nm FinFET processes. “The substrate gives us a much simpler process,” he said, adding that fewer masks should reduce process complexity and the number of opportunities to introduce defects, supporting both yield and cost.

Rather than EUV, GlobalFoundries plans to rely on advanced ArF immersion lithography, tighter overlay control, and higher-precision masks from Dresden-based mask maker AMTC.

Horstmann said the newest ASML immersion systems substantially improve overlay and critical-dimension control, while AMTC’s multi-electron-beam-written masks provide roughly 30% better overlay and CD margins than standard masks.

Physical AI needs more than logic

That process strategy matters because GlobalFoundries is not positioning FDX Fusion simply as a lower-cost route to smaller digital transistors. The platform is intended to combine dense logic with many of the analog, RF, memory, and power functions physical AI systems require.

“You can use the bulk area for high-voltage devices, and the FD-SOI area for dense logic,” Horstmann said. “That makes integrating these different functions much simpler.”

The company said FDX Fusion will combine dense digital logic with high-performance RF, analog and mixed-signal functions, low-leakage operation, embedded nonvolatile memory, and in-memory computing.

According to Horstmann, the platform should reach logic densities of roughly 14 million to 18 million gates per square millimeter while retaining the low-power and integration characteristics of the existing FDX architecture.

“You maintain the advantages of 22FDX—extremely low power and very good integration with sensors and actuators,” he said. “Extremely low power is what you need for mobile autonomous systems and drones.”

A different AI opportunity for Europe

Europe faces an uncomfortable reality in AI semiconductors: Much of the data center ecosystem is already concentrated elsewhere. “The data center AI train has left the station,” Horstmann said. “But for physical AI, the world is still open.”

His argument is that Europe does not need to win the same race. Physical AI rewards a combination of low-power compute, sensing, communications, analog, and power electronics that aligns more closely with Europe’s existing semiconductor strengths.

Dresden provides part of the industrial base for that strategy. The region already hosts GlobalFoundries, Infineon, Bosch, research organizations, suppliers, and the ESMC joint venture led by TSMC.

“The cluster is becoming self-propelling,” Horstmann said. As it grows, suppliers establish local capacity, equipment companies maintain service teams nearby, and design houses are drawn into the region.

Even ESMC, while a competitor, strengthens that ecosystem, he pointed out, by increasing Dresden’s scale and making the region more attractive to suppliers.

For Horstmann, that combination of manufacturing scale and differentiated technology creates the opening. “This is a big opportunity for Europe to define a new market,” he said, “and to be first.”

Also read:

GlobalFoundries’ Market Growth Proves U.S. Photonics Case

GlobalFoundries First to Support AI Scale-Up OCI MSA

How GlobalFoundries Takes AI from Pilot to Global Scale

AI AND BIG DATA, CHIP MANUFACTURING, FD-SOI, PHYSICAL AI

CEA-LETI, GLOBALFOUNDRIES, SOITEC