Beyond GPUs: Custom SoCs and Chiplet Architectures for Physical AI at Scale

Humanoid robots are moving from controlled demonstrations towards real-world applications. This requires more than capable AI models and advanced mechanics.

The computing architecture inside the robot must process sensor data, AI workloads and real-time control within strict limits on latency, power consumption, thermal performance, system size and cost.

The key question is therefore not simply how much computing power a robot needs, but:

"What computing architecture is appropriate for deploying Physical AI at scale?"

Matthias Neumann will discuss the role of custom System-on-Chip architectures and chiplet-based approaches in developing efficient, application-specific compute platforms for humanoid robots and other Physical AI systems.

The keynote will address how specialised computing can complement general-purpose AI processing and help connect perception, decision-making and physical control in deployable robotic systems.