✓ Next-Generation Optical Engine

Combines high-performance image processing with cutting-edge video technology.

◎ New ISP¹: Equipped with a new ISP engine. Supports Starlight ISP and AI-ISP² with AI-BNR³ for superior image quality (IQ) performance.

¹ ISP (Image Signal Processor): A semiconductor component that enhances image quality by correcting and converting image data captured by devices such as smartphones and digital cameras.

² AI-ISP: Combines AI with ISP to enable more advanced image processing than conventional ISP, improving resolution and other performance aspects.

³ AI-BNR: A feature within AI-ISP that removes image noise to produce clearer and sharper images.

◎ Encoder⁴: Supports H.264, H.265, and AV1, offering flexibility for your required formats.

⁴ Encoder: A device that converts physical motion into digital signals and outputs them.

◎ Ultra-Low Latency Streaming: Achieves near real-time video with latency under 60 ms.

◎ Color Night Vision: With regular ISP, color images can be generated at less than 1.5 Lux without using infrared LEDs (IR LEDs). Enabling AI-ISP suppresses noise even under extremely low-light conditions of 0.1 Lux⁵, delivering clear color images.

⁵ Lux: A unit that indicates illuminance, which represents the brightness of light.

✓ AI Core

Built around the 3rd-generation StarEngine IPU⁶, enabling high-performance AI processing.

⁶ IPU (Intelligent Processing Unit): A processor specialized for AI and machine learning computations.

* The 3rd generation StarEngine IPU improves upon previous generations in performance and power efficiency, supporting large-scale AI training and inference.

◎ Backward-Compatible Unified IPU SDK⁷

・Allows migration of customer AI models to the new IPU without redesign or modification.

・AI models and code developed for previous IPUs run seamlessly on the new SDK.

⁷ Unified IPU SDK: A development environment bundling essential tools such as libraries and drivers for AI development using IPUs.

◎ High Efficiency

・Optimizes hardware utilization for maximum performance.

・Enables acceleration and power savings while preserving AI model accuracy (quantization loss approximately 0.25%).

◎ Transformer⁸ Optimization

・Vision Transformer (ViT): Optimized for Vision Transformer models that deliver high performance in computer vision tasks.

・Vision Language Model (VLM): Supports processing for vision-language models that handle both visual and textual inputs.

⁸ Transformer: A foundational model used in NLP, computer vision, and multimodal AI. It is bandwidth-intensive and can be computationally heavy on SoCs without optimization.

◎ Multimodal Support

Supports AI processing that combines multiple data modalities⁹ tailored to specific applications (e.g., CLIP, Tag2Text, Grounding DINO, and MobileSAM).

⁹ Modality: The type or form of data (e.g., text, image, video).

* Supported AI Frameworks:
Broad support for major deep learning libraries and frameworks, including Caffe, PyTorch, Keras, Caffe2, TensorFlow, and ONNX.

* Evolution Across StarEngine Generations:
StarEngine versions reflect the progression of supported AI capabilities.