AMD Frames Physical AI Around End-to-End Control Timing
AMD’s October 19 physical-AI paper argues that robotics infrastructure should be measured from sensor input to physical action.
AMD’s October 19 physical-AI paper argues that robotics performance should be assessed across the full sensor-to-action loop rather than by peak compute alone.
The company separates throughput-oriented perception and reasoning from deadline-driven control, assigning different roles to GPUs, NPUs, CPUs, FPGAs and adaptive systems-on-chip. It presents distributed compute as a way to keep control functions close to sensors and actuators while isolating workloads with different safety requirements.
AMD highlights its Kria robotics platform, Versal and Zynq adaptive devices, Spartan FPGAs and open software built around ROCm and ROS 2. The paper includes AMD and commissioned third-party benchmarks; Infraday did not independently reproduce them. For infrastructure teams, the material-development signal is the push toward heterogeneous, real-time edge systems for industrial robotics.
Official source · Source reviewed October 21, 2026.