AMD unveils Kria module for real-time control, unified memory for robots

AMD has launched a robotics-focused platform built around its Ryzen AI Embedded X100 system-on-chip family and a new Kria AI system-on-module. The company is targeting high-end embedded AI and physical AI applications, including industrial arms, autonomous mobile robots, and humanoids. Its platform includes the X100 chips, Kria module, Kria AI Robotics Development Kit, open-source AMD Robotics Sophie Suite, and a Robotics Partner Network.

The X100 combines CPU, GPU, and NPU resources in a unified memory architecture. AMD said this arrangement reduces data copying and latency for perception, sensor fusion, and planning workloads compared with designs that use a CPU and discrete GPU. For firm real-time operation, it cites Linux and BIOS optimizations, with an interrupt-latency target below seven microseconds at six-nines, plus quality-of-service tools for cache reservation, memory-bandwidth allocation, and thread isolation. For hard real-time workloads, AMD recommends Zen virtualization, a FreeRTOS virtual machine, cache coloring, and virtual-machine isolation.

The Kria AI module uses the X100 and adopts the COM-HPC standard rather than a proprietary form factor. AMD said it has a 1.5-microsecond control-loop closure and industrial-oriented reliability, burn-in, test coverage, and fit-rate provisions. The development kit provides a robotics carrier card with GMSL, high-speed Ethernet, an IMU, Wi-Fi and Bluetooth, plus a baseboard FPGA for real-time I/O, sensor fusion, and safety. AMD is also making baseboard schematics available.

The system supports ROS 2 and Nav2 acceleration and AMD ROCm AI software. AMD claimed performance advantages over NVIDIA Thor in selected system-level and real-time measures. Early users named by AMD include Castec International and Foundation Robotics, while listed partners include Open Robotics, Open Navigation, and OpenCV.