CPU+MCU+NPU Collaborative Architecture Enables “Perception-Decision-Action” Closed Loop
The RDK S100 adopts a brain-cerebellum architecture, integrating brain and cerebellum functional modules into a single SoC platform. This not only reduces hardware costs but also enhances overall performance. It supports efficient collaboration between large and small embodied intelligence models and provides rich peripheral interfaces, as well as full-link development infrastructure support for software-hardware coordination and edge-cloud integration.
For the brain-cerebellum structure of hierarchical decision-making models, a single SoC platform that balances low power consumption and diverse computing power combinations is required. The RDK S100 integrates CPU+BPU+MCU on a single SoC to achieve a trinity of “perception + decision + action”:
- Brain: 6-core Arm Cortex-A78AE CPU (supports real-time kernels, featuring efficient scheduling and low latency) + 80TOPS BPU (optimized for CNN and Transformer, enabling faster inference and lower power consumption), meeting computational needs for complex decision-making and planning tasks.
- Cerebellum: 4-core Arm Cortex-R52+MCU, providing robots with high-frame-rate, low-latency real-time joint control capabilities for various real-time motion control scenarios.
This architecture allows the RDK S100 to support dynamic fusion and seamless switching of large and small models. On one hand, it can deploy popular models such as visual/point cloud detection, LLM, and VLM to provide accurate data support for robot decision-making. On the other hand, it can deploy multiple motion control models to enable stable and reliable high-speed, high-precision motion control. Meanwhile, the close collaboration between MCU and BPU reduces CPU load by 80% and improves robot response speed.
Over 50+ Customers Evaluated, Applied to Humanoid/Quadruped Robots and More
Currently, D-Robotics has partnered with over 20+ leading embodied intelligence customers. The RDK S100 has been evaluated by more than 50+ customers and applied to various products, including:
– Bipedal/point-foot robots, small biped robots, semi-humanoid robots, quadruped robots, LeRobot mechanical arms, humanoid robots, BEV detection, multi-channel video detection systems
At the launch event, the RDK S100 was demonstrated in Unitree G1 robots, Reeman robots, and multiple mechanical arms performing tasks such as dance performances, object sorting, and complex terrain traversal.
RDK S100 Technical Specifications
- AI Performance – 80 TOPS
- CPU – 6x Arm [email protected]
- MCU – 4x Arm [email protected]
- GPU – 1x Arm Mali-G78AE@100 GFLOPS
- Memory – 12GB LPDDR5, 96-bit
- Storage
- 64GB eMMC
- M.2 Key M (PCIe 3.0 x1) * 1
- Display – HDMI 1.4 up to 2K@60FPS
- Network
- Gigabit Ethernet (RJ45 Connector) * 2
- M.2 Key E (PCIe 3.0 x1, UART) Connector for Wi-Fi & Bluetooth
- USB
- USB 2.0 (Type-C) * 1 for flash & debug
- USB 3.0 * 4
- Expansion Header
- JTAG Header * 1
- Automatic Header * 1
- 40-pin Header for Main GPIO * 1
- 16-pin Header for MCU GPIO * 1
- Camera Expansion Header * 1
- MCU Port Expansion Header * 1
- Power – 12-20V DC
- Mechanical – 120x121x51mm (Case Included)
Half the Price of NVIDIA Products with Similar Computing Power
These features are critical for embodied robot developers. Equally important is the pricing: the more powerful RDK S100P with larger memory will launch in Q3 this year, while the RDK S100 is now available at a limited-time discount of 389USD — nearly half the cost of NVIDIA’s product with similar computing power. At this price, the RDK S100 truly aims to “be friends with embodied robot developers.”
For international customers, AnalogLamb.Comsupports pre-order purchases, send email to us.
