Physical AI Specialized Semiconductor Platforms for the Next Wave of Automation

Related Vendors

FAULHABER Drive Systems

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

Wuxi InfiMotion Technology Co., Ltd.

Physical AI marks the transition from AI that primarily runs in data centers to intelligent systems that operate directly in the physical world. In technical terms, this involves machines that sense, think, act and communicate (STAC) with other systems, interpreting and reacting to their environment in real time.

The great technological leaps of recent years have been achieved in AI data centers, which today form the basis for training and hosting large models. However, the next step in value creation is taking place at the periphery: in billions of networked endpoints that bring intelligence directly into products and processes. Physical AI shifts computing and decision-making logic closer to the place where data is generated and actions are triggered - in vehicles, robots, machines and devices. This fundamentally changes the requirements: the decisive factor is no longer just maximum computing power, but the ability to work reliably in real time under tight energy, latency and cost budgets. At the same time, systems must remain updatable over many years in order to load new models and functions without replacing the hardware.

STAC: The Real-Time Operating Model of Physical AI

A practical framework model for physical AI is the STAC approach (sense - think - act - communicate), which describes how such systems work in the field.

• Sense: Recording of environmental data using multimodal sensor technology - such as cameras, radar, lidar, audio, haptics and environmental sensors.
• Think: Local processing and interpretation of this data, often with AI inferencing, to make deterministic, reliable decisions in real time.
• Act: Conversion of these decisions into precise movements or switching operations via motors, actuators and control loops with tight feedback loops.
• Communicate: Exchange of status, sensor data and models between device, edge and cloud to orchestrate fleets, roll out updates and continuously optimize systems.

From a semiconductor perspective, STAC derives three key requirements: high energy efficiency, deterministic latency and long-term reliability including security over the entire life cycle. Only if all four STAC functions can be mapped under these boundary conditions can physical AI applications be transferred from the pilot phase to safety-critical, scaled environments.

Convenient Semiconductor Platforms for Physical AI

The workloads described above mean that classic SoCs optimized purely for maximum computing power are reaching their limits. What is needed are purpose-built platforms that combine compute, sensing, actuation, memory and connectivity under a common energy and latency budget. Globalfoundries, for example, can address this need with several technology modules:

FDX and FinFET:

• Ultra-low power CMOS with body bias capabilities and efficient FinFET nodes that combine high compute density with very low power consumption.
• Integration of analog, mixed-signal and RF functionality on one die, enabling compact edge SoCs with integrated sensor connectivity.

Embedded non-volatile memory (eNVM):

• Local storage of firmware, models and configuration data, allowing systems to be adapted in the field for years via OTA updates.
• Basis for software-defined architectures that use hardware platforms for multiple generations of software and AI models.

Power Management Technologies:

• Fine-grained voltage and clock domains, deep sleep states, and adaptive controllers to minimize both active and standby losses. 
• This is particularly relevant in battery-operated or passively cooled systems such as robotics, wearables, or connected sensors.

Silicon Photonics and RF:

• High data rates and bandwidths for on-board and off-board connections to transmit large volumes of sensor data with minimal energy expenditure. 
• Scalable connectivity across billions of devices is a fundamental requirement for distributed physical AI fleets.

Advanced Packaging and Heterogeneous Integration:

• 2.5D and 3D approaches that integrate compute, memory, RF, and power chips into a logical system. 
• Optimized signal paths and energy flows to reduce latencies while simultaneously minimizing footprint and unit costs.

These building blocks enable physical AI systems to achieve more with less—meaning they can handle demanding STAC workloads within tight power, thermal, and compute limits. This opens up the possibility for the industry to bring intelligent behavior to a significantly broader range of device classes than would be economically feasible with purely high-end SoCs.

Software-Defined, Distributed Intelligence

A second structural transformation pertains to system architecture: instead of centralized compute blocks, distributed, software-defined systems are emerging. Intelligence is migrating closer to sensors and actuators to make decisions where they are needed—using short signal paths and reproducible latency. For semiconductors, this means:

• Devices must cover broad, heterogeneous workloads—from classic control and signal processing to AI inferencing and safety functions.
• Platforms are differentiated through software; hardware is deliberately designed to be generic enough to outlast multiple product cycles.
• Embedded memory, secure boot and update mechanisms, and hardware security features become mandatory, not optional.

GlobalFoundries, for example, explicitly positions its platforms for such software-defined, distributed architectures and couples process technologies with IP, packaging, and manufacturing in a co-design approach. By integrating the MIPS portfolio—including real-time capable, multi-threaded RISC-V processors—the compute building blocks are supplemented to ensure deterministic real-time capability in such systems.

Application Areas: ADAS, Robotics, Humanoids

The impact of these technical developments is already evident in several segments:

• Automotive / ADAS: Physical AI systems drive the next generation of Advanced Driver Assistance Systems and automated driving functions. They combine multimodal sensor technology (cameras, radar, lidar) with real-time processing and precise actuation to execute driving maneuvers under strict safety, power, and thermal requirements.
• Industrial Automation and Logistics: Autonomous mobile robots, cobots, and driverless transport systems must navigate dynamic environments, recognize objects, plan paths, and interact safely with humans. Here, energy-efficient edge platforms with integrated sensor technology, connectivity, and secure control pay off.
• Humanoids and Complex Robotics: For humanoid robots intended to work in human environments, high sensor density, distributed intelligence, deterministic latency, and robust communication are indispensable. Market forecasts suggest a very high long-term potential in this segment alone, further underscoring the relevance of scalable physical AI platforms.

With the transition from cloud-centric AI to physical AI, the focus is shifting towards reliable, energy-efficient, and adaptable systems in the field. Semiconductor platforms that holistically address STAC workloads while considering power, latency, connectivity, and updatability are becoming critical enablers for the next wave of automation. GlobalFoundries aims to fulfill this role with differentiated process technologies, advanced packaging, RF and connectivity solutions, and the integration of MIPS processor IP. For developers in automotive, industry, and robotics, this creates an ecosystem that allows physical AI systems to transition from initial demonstrators to widely deployed, critical applications. (sg)

*Ed Kaste is Senior Vice President of Ultra-Low Power CMOS Business at Globalfoundries.