Vimicro Unveils Single-Chip DeepSeek LLM Solution

Revolutionizing AI Chips: Zhongxing Micro Unveils Single-Chip DeepSeek Large Model Capabilities

Zhongxing Microelectronics, also known as Vimicro, has recently marked a significant advancement in the field of artificial intelligence with the introduction of its newest AI chip, the ‘Starlight Intelligence No. 5.’ This state-of-the-art chip was presented at the 8th Digital China Construction Summit in Fuzhou on April 30th. The chip’s standout feature is its capability to autonomously run DeepSeek’s 7B, 8B, and 16B large models, eliminating the need for external processing capabilities. This achievement positions it as the foremost fully autonomous and controllable embedded AI chip capable of simultaneously running both general-purpose language models and visual large models on a single chip.

Groundbreaking Architecture and Enhanced Performance

The ‘Starlight Intelligence No. 5’ chip is built around Zhongxing Micro’s proprietary general-purpose multi-core heterogeneous GP-XPU architecture. This innovative design offers significant enhancements over traditional CPU+GPU architectures in terms of operational efficiency, real-time performance, cost-effectiveness, and security. The GP-XPU architecture is uniquely optimized for AI workloads, enabling the chip to manage complex tasks with enhanced speed and precision.

Key Advantages of the GP-XPU Architecture:

• Enhanced Efficiency: The architecture maximizes the use of computing resources, resulting in quicker processing times and superior overall performance.
• Real-time Performance: The chip can handle real-time data processing, making it perfect for applications demanding immediate responses.
• Cost-Effectiveness: The chip’s design reduces the need for costly external components, offering a more affordable solution.
• Security: The architecture incorporates security features that protect against unauthorized access and data breaches.

Domestically Produced and Fully Controllable

The ‘Starlight Intelligence No. 5’ chip is based on domestically produced process technology, guaranteeing total autonomy and control over its production and application. This is particularly crucial for applications in strategically important sectors where reliance on foreign technology can pose potential risks. By utilizing domestically sourced components and manufacturing processes, Zhongxing Micro aims to decrease dependency on external suppliers and strengthen China’s standing in the global AI market.

Applications in Key Sectors:

• National Strategic Importance: The chip can be deployed in critical infrastructure and defense systems, providing a secure and reliable AI processing solution.
• Urban Perception: The chip can be used in smart city initiatives to analyze data from sensors and cameras, enabling real-time monitoring and management of urban environments.
• Intelligent Manufacturing: The chip can enhance automation and efficiency in manufacturing processes, improving product quality and reducing costs.
• Smart Agriculture: The chip can be used to analyze data from agricultural sensors and drones, optimizing crop yields and resource utilization.
• Intelligent Transportation: The chip can power advanced driver-assistance systems (ADAS) and autonomous vehicles, improving safety and efficiency on the roads.

Breaking the Edge Computing Barrier

The ‘Starlight Intelligence No. 5’ chip is designed to overcome the limitations of edge computing, enabling AI processing to be performed directly on devices without relying on cloud-based resources. This is made possible through the chip’s advanced architecture and its ability to run large AI models on a single chip.

GP-XPU Architecture Components:

• RISC-V CPU: A high-performance RISC-V CPU delivers the general-purpose processing power needed to handle a wide range of tasks.
• GP-GPU: A general-purpose GPU accelerates AI workloads, enabling faster processing of complex algorithms.
• NPU: A neural processing unit (NPU) is specifically designed for AI inference, optimizing the performance of neural networks.
• ISP: An image signal processor (ISP) enhances the quality of images captured by cameras, improving the accuracy of visual AI applications.
• VPU: A video processing unit (VPU) accelerates video encoding and decoding, enabling efficient video streaming and analysis.
• ECU: An encryption/decryption unit (ECU) provides hardware-based security for sensitive data.

Heterogeneous Computing Pool (HCP)

The chip utilizes a dedicated Heterogeneous Computing Pool (HCP) task scheduling unit and a secure memory management system to enable real-time scheduling and dynamic sharing of computing and storage resources among the various heterogeneous cores. This significantly improves computing efficiency and data throughput, allowing the chip to deploy and run DeepSeek 1.5B, 7B, 8B, and 16B general-purpose large models and visual large models on a single chip. The innovative HCP dynamically allocates resources, ensuring that each processing unit receives the necessary support in real-time, significantly optimizing overall performance.

This sophisticated resource management system is crucial for efficiently running large AI models. The ability to dynamically adjust resource allocation based on the real-time demands of different tasks optimizes processing speed and reduces latency, making it ideal for applications requiring instant responses and complex calculations.

Edge Computing Capabilities:

• Real-time Video Analysis: The chip can perform real-time detection, recognition, and tracking of objects in video streams.
• Natural Language Processing: The chip can process natural language inputs, enabling applications such as chatbots and voice assistants.
• Task Planning: The chip can plan and execute complex tasks, making it suitable for robotics and automation applications.
• Knowledge Management: The chip can store and retrieve information from knowledge bases, enabling intelligent decision-making.
• Automatic Control: The chip can control actuators and other devices, enabling closed-loop control systems.

By deploying eight ‘Starlight Intelligence No. 5’ chips in conjunction, it is possible to support the operation of ‘full-blood’ 671B parameter DeepSeek large models and visual large models, completely breaking the bottleneck of computing power at the edge. This parallel processing capability significantly expands the range of applications that can be supported at the edge, from complex simulations to real-time data analysis. The coordinated operation of multiple chips enables unparalleled performance, making previously impossible edge-based AI solutions a reality.

A Versatile AI Chip

The ‘Starlight Intelligence No. 5’ chip is a versatile AI chip that excels in a variety of applications. Its key advantages include high efficiency, low energy consumption, low cost, small size, strong security, and high adaptability. The combination of these attributes makes it an ideal solution for deploying AI in diverse and challenging environments.

Key Advantages in Detail:

• High Efficiency: The chip utilizes an operator-level MoE architecture and HCP real-time scheduling mechanism to improve computing efficiency by approximately 40% and data throughput by approximately 50%. This means that more tasks can be performed with less resources, increasing operational output and effectiveness.
• Low Energy Consumption: The chip’s heterogeneous computing resources are allocated on demand, reducing energy consumption by at least 30%. This lowers operational costs and environmental impact, making it a sustainable solution for long-term deployments.
• Low Cost: The chip’s ‘end-edge closed-loop architecture’ reduces reliance on cloud resources, resulting in a comprehensive deployment cost that is approximately 1/3 of that of a server architecture with the same performance. This affordability makes advanced AI technology accessible to a wider range of applications and industries.
• Small Size: The processing board embedded with the chip requires a space of only about the size of a business card, which can meet the space requirements of various miniaturized intelligent devices. This allows for seamless integration into existing systems and supports the development of innovative new devices.
• Strong Security: The chip uses memory partition security isolation technology, supports high-performance domestic encryption algorithm calculations, achieves source-level password protection, and has native data security features. This robust security framework protects sensitive data from unauthorized access and cyber threats, ensuring data integrity and confidentiality.
• High Adaptability: The chip supports open ecosystems and mainstream AI frameworks, seamlessly connecting to various application scenarios in thousands of industries. In terms of working environment adaptability, the ‘Starlight Intelligence No. 5’ chip eliminates the dependence of cloud-based large computing equipment on the constant temperature and humidity environment of the computer room, and can complete high-performance intelligent work in various production and life environments. This flexibility makes it a versatile solution suitable for a broad spectrum of environments and applications.

Empowering ‘Localized Intelligent Agents’

The ‘Starlight Intelligence No. 5’ chip is designed to empower ‘localized intelligent agents,’ which are AI-powered systems that can operate independently on devices without relying on cloud-based resources. This approach offers several advantages, including improved response times, reduced costs, and enhanced security. These localized agents can perform complex tasks autonomously, providing immediate and reliable solutions in various settings.

Advantages of Localized Intelligent Agents:

• Improved Response Times: By processing data locally, the chip can respond to events in real-time, without the delay associated with cloud-based processing. This is crucial for applications where immediate action is required, such as emergency response systems or real-time control applications.
• Reduced Costs: By reducing reliance on cloud resources, the chip can significantly lower the cost of AI deployments. This makes advanced AI technology more accessible and affordable for a wider range of applications and industries.
• Enhanced Security: By processing data locally, the chip can protect sensitive information from unauthorized access and data breaches. This is particularly important for applications dealing with personal or confidential data, ensuring privacy and security.

Applications of Localized Intelligent Agents:

• Smart Homes: Localized intelligent agents can control smart home devices, such as lights, thermostats, and appliances, providing a more convenient and efficient living experience. This includes personalized control, automated routines, and energy management, improving overall comfort and efficiency.
• Wearable Devices: Localized intelligent agents can analyze data from wearable sensors, providing personalized health and fitness recommendations. This enables proactive health management, customized workout plans, and real-time feedback, promoting a healthier lifestyle.
• Industrial Automation: Localized intelligent agents can control robots and other industrial equipment, improving efficiency and safety in manufacturing environments. This includes automated quality control, predictive maintenance, and real-time process optimization, enhancing productivity and reducing downtime.

Reconstructing the AI Deployment Paradigm

Zhongxing Micro’s ‘Starlight Intelligence No. 5’ chip is poised to transform the AI deployment paradigm by enabling a more distributed and localized approach. This will reduce reliance on cloud-based resources and empower a wider range of applications. The shift from centralized, cloud-dependent AI systems to distributed, edge-based solutions marks a significant advancement in the field.

Key Benefits of the New Paradigm:

• Reduced Latency: Localized processing eliminates the need to transmit data to the cloud, reducing latency and improving response times. This makes AI applications more responsive and effective, especially in time-sensitive situations.
• Lower Bandwidth Requirements: Localized processing reduces the amount of data that needs to be transmitted over the network, lowering bandwidth requirements and costs. This is particularly beneficial in environments with limited or expensive network connectivity.
• Enhanced Privacy: Localized processing keeps sensitive data on the device, protecting it from unauthorized access and data breaches. This ensures data privacy and confidentiality, which is crucial for maintaining user trust and compliance with privacy regulations.
• Improved Reliability: Localized processing allows AI systems to continue operating even when the network connection is interrupted. This ensures uninterrupted service and resilience in challenging environments, improving overall reliability and performance.

According to Zhang Yundong, CEO of Zhongxing Micro Technology, the ‘Starlight Intelligence No. 5’ will promote the intelligent upgrading of the end side and the edge side, greatly reduce the dependence on cloud computing resources, save system construction costs, and explore an artificial intelligence technology route that conforms to China’s national conditions. This vision emphasizes the importance of developing AI technologies that are tailored to the specific needs and conditions of the local environment.

The traditional AI reasoning process that relies on the cloud often faces problems such as large response delays, high costs, and restrictions on the network environment. The chip enables large model reasoning to be directly closed-loop on the edge side through localized computing power and endogenous security mechanisms, which not only saves cloud transmission and storage costs, but also greatly improves the response speed, achieving millisecond-level response, and protects data throughout its lifecycle with a source-level password mechanism. This localized approach addresses the limitations of cloud-based AI, offering improved performance, security, and cost-effectiveness.

This technological breakthrough provides a comprehensive optimization of efficiency, energy consumption, cost, volume, security, and adaptability for scenarios such as urban perception, intelligent manufacturing, and intelligent transportation, truly enabling thousands of industries to create autonomous and controllable intelligent terminals. The chip’s ability to optimize these factors simultaneously makes it a game-changing solution for deploying AI in various applications.

Zhang Yundong stated: ‘In response to the actual needs of digital transformation in China’s traditional industries, Zhongxing Micro is developing distributed artificial intelligence computing technology for the end, edge, and cloud. Without relying on the most advanced semiconductor process technology, we insist on original innovation and independent control, and explore an artificial intelligence technology route that conforms to China’s national conditions.’ This statement underscores the company’s commitment to innovation and independent development, aiming to create AI solutions that meet the unique challenges and requirements of China’s digital transformation. The focus on distributed computing technology allows for a more flexible and scalable approach to AI deployment, enabling a broader range of applications across different sectors.