Embedded AI Accelerators & Vision Pipelines

Running computer vision, depth perception, and neural inference models at the physical edge requires direct access to specialized silicon: GPUs, NPUs, and hardware video encoders.

Admiral provides out-of-the-box hardware integration for dedicated AI accelerators across x86_64, NVIDIA Jetson, and ARM64 platforms, eliminating manual driver installation and container driver bloat.

---

1. NVIDIA Jetson & Desktop GPU Support

On machines equipped with NVIDIA hardware (including x86 GPUs today, with NVIDIA Jetson platforms such as AGX Orin, Orin NX, and Orin Nano on the active roadmap), Admiral OS automatically detects the GPU hardware and injects necessary kernel device nodes into container workloads:

/dev/nvidia0          # Primary GPU execution device
/dev/nvidiactl        # NVIDIA control device
/dev/nvidia-uvm       # Unified memory management device
/dev/nvidia-uvm-tools # Profiling and diagnostic interface
/dev/nvhost-*         # Jetson hardware accelerators (VIC, NVENC, NVDEC)

• No Container Driver Bloat: Container images do not need to bundle heavy host drivers. Workloads utilizing CUDA, TensorRT, or ROS perception stacks run directly against host-injected device nodes.
• Unified Memory (/dev/nvidia-uvm): Zero-copy shared memory transfers between system RAM and GPU VRAM, critical for low-latency point-cloud processing and camera frame inference.

---

2. Rockchip RK3588 NPU Acceleration

For low-power autonomous systems, smart cameras, and drone companion computers operating under strict 15W thermal budgets, Admiral provides native pass-through for the Rockchip RK3588 6 TOPS NPU:

• Direct /dev/rknpu Node: The Rockchip Neural Processing Unit is mounted directly into container workloads with full read/write permissions.
• RKNN-Toolkit Support: Run INT8 and FP16 quantized YOLOv8, MobileNet, and custom vision models with sub-10ms inference times at minimal power draw.

---

3. High-Throughput Camera & Vision Ingestion

Vision-guided robots and AI perception nodes ingest massive visual data streams:

Interface Standard	Linux Device Node	Application Usage
USB 3.0 RealSense / OAK-D	/dev/video*, /dev/bus/usb	Intel RealSense D435/D455, Luxonis OAK-D stereoscopic cameras with raw depth maps and IMU sync.
Industrial MIPI CSI-2	/dev/video*, /dev/media*	Direct sensor deserialization (GMSL2 / FPD-Link III) for high-bandwidth automotive and drone cameras.
Network IP Cameras	Host Network UDP/RTSP	Multi-camera RTSP and WebRTC video decoding via hardware-accelerated NVDEC pipelines.

---

4. Zero-Copy Shared Memory Frame Pipelines (/dev/shm)

High-resolution stereoscopic camera feeds and 3D LiDAR point clouds generate gigabytes of raw data per second. Passing these buffers between independent container processes over standard network sockets creates severe CPU serialization bottlenecks.

• POSIX Shared Memory: By configuring a sized tmpfs mount on /dev/shm (e.g. 2GB to 8GB), sensor acquisition drivers write raw camera frames directly into shared memory.
• Zero-Copy Loaning: Perception and SLAM nodes (via FastDDS SHM or Iceoryx) read the raw memory pointers directly with zero memory copy overhead, leaving physical CPU cores free for path planning and navigation control.