Specifications
AI Performance
80 TOPS (S100) / 128 TOPS (S100P) on-device inference
SoC Architecture
D-Robotics single-SoC computation-control integrated platform ('brain-cerebellum' architecture)
CPU
6x Arm Cortex-A78AE @ up to 1.5 GHz, up to 2.0 GHz on the S100P
Real-time MCU
4x Arm Cortex-R52+ @ 1.2 GHz for low-latency, high-frame-rate motion control
GPU
1x Arm Mali-G78AE @ approximately 100 GFLOPS
NPU
1x BPU Nash core supporting 160+ ONNX operators, optimised for CNN and Transformer models
Memory
12 GB or 24 GB LPDDR5, 96-bit
Storage
64 GB eMMC with pre-installed system, plus M.2 Key-M (PCIe 3.0 x1) for SSD expansion
Display
HDMI 1.4 up to 2K@60FPS
Networking
2x Gigabit Ethernet RJ45; M.2 Key-E (PCIe 3.0 x1, UART) for Wi-Fi and Bluetooth
USB
4x USB 3.0 plus 1x USB 2.0 Type-C for flashing and debug
Camera Inputs
Dual MIPI camera inputs, with GMSL options available through expansion boards
Expansion
2x PCIe 3.0 interfaces plus MCU peripheral expansion headers
LLM Support
DeepSeek-R1-Distill-Qwen 1.5B/7B, Qwen2.5 1.5B/7B and Qwen2.5-Omni 3B via the D-Robotics LLM toolchain
Measured LLM Throughput
Qwen2.5-1.5B q8 on S100P: 130 ms TTFT at 24.0 tokens/s. DeepSeek-R1-Distill-Qwen-1.5B: 109 ms TTFT at 27.1 tokens/s
Software
RDK robot OS (TROS) with ROS 2 support and a full software-hardware co-development toolchain
Target Robots
Humanoids, quadrupeds, wheeled robots, robotic arms and autonomous mobile platforms
Demonstrated On
Unitree G1 robots, Reeman robots and multiple robotic arms performing dance, sorting and complex terrain traversal
Overview
The RDK S100 is the first robot development kit to integrate computation and real-time control into a single SoC. Rather than pairing a compute module with a separate motion-control board, it combines a 6-core Cortex-A78AE application cluster, a Nash BPU for AI inference and a 4-core Cortex-R52+ MCU cluster on one platform — what D-Robotics calls a brain-cerebellum architecture.
The split of duties is deliberate: the A78AE cores handle multi-task logic, coordination and real-time scheduling; the Nash BPU runs large-model inference with support for 160+ ONNX operators and optimisation for both CNN and Transformer networks; and the R52+ cluster delivers the low-latency, high-frame-rate control loops that robot actuators demand. Memory is 12 GB or 24 GB of LPDDR5 with 64 GB of eMMC and an M.2 slot for SSD expansion.
On-device LLM performance is documented rather than theoretical: the D-Robotics toolchain reports 130 ms time-to-first-token at 24.0 tokens per second for Qwen2.5-1.5B q8, and 109 ms TTFT at 27.1 tokens/s for DeepSeek-R1-Distill-Qwen-1.5B on the S100P. The platform ships with RDK robot OS and ROS 2 support, targeting humanoids, quadrupeds, wheeled robots and robotic arms. QS Compute supplies embedded AI development kits, robotics compute modules and edge AI systems — request a quote.
Key Benefits
Compute and control in one SoC: removes the separate motion-control board. Real-time MCU cluster: 4x Cortex-R52+ for actuator-rate control loops. Transformer-ready NPU: Nash BPU optimised for large models, 160+ ONNX operators. Measured LLM serving: documented TTFT and token rates for quantised 1.5B-class models.
Applications
Humanoid and quadruped robot development, robotic arms and pick-and-place, autonomous mobile robots, embodied intelligence research, vision-guided manipulation, and edge LLM/VLM deployment on robots.
Request a Quote — D-ROBOTICS RDK S100 — 80/128 TOPS ROBOT DEVELOPMENT KIT
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →