AAEON Launches Open Robotics Development System for Humanoids

AAEON released the CEXD-INTRBL, an open robotics development system built around the Intel Core Ultra X7 358H SoC. The platform combines a 16-core CPU, Arc B390 GPU, and an on-chip NPU to deliver 180 TOPS of aggregate AI acceleration, positioning the unit as a compact, self-contained edge computer for humanoid robots, autonomous mobile robots, and vehicle prototyping. The system ships with onboard LPDDR5 up to 64GB, a default 256GB NVMe, and supports Windows 11 and Ubuntu 24.04/25.04.

The CEXD-INTRBL is engineered for multi-sensor robotics workloads and real-time control. Key hardware and system specs include:

• •SoC: Intel Core Ultra X7 358H (16 cores in 4P+8E+4LPE configuration, up to 4.8 GHz), 18 MB cache
• •Graphics and AI: Arc B390 12-core GPU (up to 2.5 GHz, ~122 TOPS) plus a dedicated NPU (~50 TOPS) for a combined 180 TOPS AI capability
• •Memory and storage: Up to 64GB LPDDR5 8533MT/s onboard, M.2 2280 M-Key NVMe (256GB default) and additional M.2 expansion
• •I/O and sensors: 2x FAKRA connectors supporting up to eight GMSL2 cameras, 2x MIPI CSI internal, 4x USB 3.2 Gen1 Type-C, 2x USB 2.0, and audio
• •Networking and synchronization: 4x 2.5GbE via Intel I226-V controllers with IEEE1588 PTP support for sub-millisecond sync across sensors and actuators
• •Control and fieldbus: CAN Bus, 2x RS-232, 2x RS-485 and a 40-pin HAT with 22 GPIOs for motion controllers and low-level peripherals
• •Power and environment: 19V-24V DC input, 200W-250W consumption nominal, operating range 0°C to 50°C with airflow

The combination of high on-chip TOPS, deterministic networking, and extensive sensor I/O addresses three recurring engineering bottlenecks in robotics: compute locality, time-synchronized multi-sensor fusion, and integration complexity. By consolidating CPU, GPU, and NPU on a single SoC and pairing it with IEEE1588 PTP-capable 2.5GbE ports and GMSL camera inputs, the CEXD-INTRBL reduces the need for external accelerators, separate synchronization hardware, and complex prototyping boards.

Developers can run inference pipelines locally using Intel toolchains such as OpenVINO, prototype sensor fusion across up to eight cameras, and exercise real-time motion stacks over CAN Bus without adding gateway devices. Onboard LPDDR5 at 8533MT/s and the PCIe-capable M.2 sockets enable higher-throughput models and faster data logging. The unit supports standard desktop OSes, which simplifies deployment of existing ROS, perception, and planning stacks.

This release follows a broader industry shift toward heterogeneous SoC designs that embed NPUs alongside CPUs and GPUs, enabling edge-first AI that minimizes round-trip latency to the cloud. AAEON targets a developer audience that needs compact, ruggedized compute with robotics-specific connectors like GMSL and CAN. Competing approaches often split compute across discrete GPUs, FPGAs, or separate NPUs. The CEXD-INTRBL opts for a single integrated SoC strategy that favors power efficiency and system simplicity over absolute peak throughput of a server GPU.

Evaluate real-world performance on multi-camera perception pipelines and closed-loop control tasks to confirm the claimed 180 TOPS translates to practical latency gains. Pay attention to thermal behavior under sustained loads and ecosystem support from Intel toolchains and community ROS packages. If the platform delivers on synchronized, high-throughput sensor fusion with low power, it will accelerate prototyping cycles for humanoid and autonomous systems.