- Quad Arm Cortex®
- A53 @1.45 GHz
- Neural Process Unit
- 8 TOPS @1.45 GHz
AI-G
Edge AI-focused
Single Board Computer
Overview of
N-Dolphin
Built on Telechips’ energy-efficient AI SoC, the TOPST AI-G board delivers 8 TOPS of AI computer
through a dual-cluster NPU — optimized for real-time, on-device inference in vision-based applications.
It excels in tasks like object detection, facial recognition,
and smart surveillance with low latency and high efficiency at the edge.
Powered by a Cortex-A53 quad-core CPU and equipped with MIPI CSI/DSI, PCIe 3.0, Gigabit Ethernet,
and a 40-pin GPIO header, it offers broad integration flexibility.
Its compact design, low-power operation, and support for Linux and leading AI frameworks make it ideal
for deployment in smart city, mobility, and industrial AI systems.
Improved compatibility and usability
with more flexible interface support.
Efficiently Develop On-device AI solution
AI-G provides solutions for developing vision processing-based on-device AI applications. Specialized in CNN algorithms, it is adaptable for use in ADAS/autonomous systems, robots, and IoT devices with cameras. With our dedicated AI SDK, developers can run a range of popular AI frameworks on AI-G, ensuring compatibility, performance, and scalability for edge AI innovation.
2 Cluster
Two separate clusters can each have an independent AI system. With two parallel 4 TOPS clusters, AI-G enables more stable and scalable inference performance, ideal for running multiple AI models or distributing workloads efficiently.
ML Framework
Compatible with major machine learning frameworks, the platform allows efficient development and deployment of various AI models. This flexibility makes it easy to integrate with existing AI pipelines and accelerate time-to-market for edge applications.
Low Power Consumption
Optimized for the demands of edge devices where power efficiency is critical, AI-G delivers up to 2.5 TOPS per watt, making it ideal for always-on applications such as smart cameras, autonomous robots, and IoT sensors. Its energy-efficient NPU architecture ensures sustained AI performance without thermal or power overhead.
AI Toolkit
An AI toolkit for deploying GPU-trained neural networks on NPU-based hardware. It supports INT8 quantization, optimization, compilation, and simulation to enable efficient model conversion and accurate performance evaluation for edge inference.
Camera Interface
Optimized for real-time vision AI, the platform supports 4-lane MIPI CSI input with a 4-channel ISP and RGB-IR sensor compatibility. Advanced features like 2DNR/3DNR and de-warping ensure clean input for edge inference. Ideal for camera-based applications such as object detection and real-time recognition.
Supported AI Operators
Supports a focused set of core AI operators, optimized for typical CNN-based applications. Ensures stable and efficient inference in key use cases such as object detection and facial recognition.
| Category |  |  |  |
|---|---|---|---|
| Convolution |
|
|
|
| Activation Function |
|
|
|
| Pooling |
|
|
|
| Element-wise Ops |
|
|
|
| Normalization |
|
| |
| Sampling & Resizing |
|
|
|
| Reduction |
| ||
| Tensor Operations |
|
|
|
| Shape Manipulation |
|
|
Specification
| SoC (N-dolphin) |
|
|---|---|
| Memory |
|
| eMMC |
|
| Video Out |
|
| Camera In |
|
| PCIe |
|
| Ethernet |
|
| General Function Interface |
|
| Vehicle Signals Interface |
|
| Debug |
|
| Power |
|
| OS |
|
| PCB |
|