Ruigu Electronic

As we enter 2025, automotive companies have recognized that the demand for high-performance chips has become a key driver of industry progress under the end-to-end single-stage model. NVIDIA’s Thor chip, launched in 2022, is hailed as a new benchmark in the field of intelligent driving, bringing new possibilities to the automotive industry with its technical performance and design philosophy.

This article provides an in-depth technical analysis of the Thor chip, exploring its architectural features, computational capabilities, system integration, and future technological trends.

Architectural Innovation: Based on the Blackwell Architecture

The Thor chip is based on NVIDIA’s latest Blackwell architecture, which reflects the company’s deep expertise in high-performance computing.

● Core features of the Blackwell architecture include:

◎ Advanced manufacturing process: The Thor chip is manufactured using TSMC’s 4-nanometer process, paired with CoWoS-R packaging technology. This advanced process significantly enhances the chip’s transistor density and energy efficiency, providing a solid foundation for high computing power and low power consumption.

◎ Modular design: The Blackwell architecture adopts a modular design, enabling flexible configuration of compute units and memory bandwidth. The Thor chip offers five different configurations to meet the diverse needs of systems ranging from entry-level to high-end intelligent driving systems.

◎ Dedicated compute cores: The Thor chip integrates multiple dedicated compute cores, including Tensor Cores for AI inference, CUDA cores for graphics processing, and dedicated accelerators for neural network training and optimization. These cores work together to significantly enhance the chip’s overall performance.

The top-of-the-line Thor-Super model boasts a computational capability of up to 2,000 TOPS (trillion operations per second), nearly eight times that of its predecessor, the Orin chip. The Thor chip can run end-to-end (E2E) autonomous driving models in real time, which require processing large amounts of data and performing complex computations.

The Thor chip can simultaneously process data from multiple sensors, including cameras, radar, and lidar, enabling multimodal perception and decision-making. The launch of the Thor chip marks the industry’s transition to the Intelligent Driving 3.0 era, characterized by a dual-system architecture combining “end-to-end + VLA.”

VLA (Visual Language Action Large Model) serves as a new-generation technological framework, offering more intelligent and flexible solutions in complex driving scenarios.

● The key role of VLA: Enhancing the autonomous driving system’s understanding of scenarios and improving the system’s decision-making accuracy.

● The core role of Thor: Its high computing power supports the deployment of the full VLA model on the vehicle end, addressing the current computing power bottleneck of the Orin-X platform.

The application of the NVIDIA Thor chip at this stage not only meets the current needs of autonomous driving development but also lays the foundation for higher-level autonomous driving technologies.

One of the most disruptive innovations of the Thor chip lies in its system integration capabilities. Automotive electronic systems are typically divided into multiple controllers, each handling ADAS (Advanced Driver Assistance Systems) and IVI (In-Vehicle Infotainment Systems). This distributed architecture suffers from data latency and reliability issues. The Thor chip adopts a single-chip design, unifying ADAS and IVI systems onto a single platform.

● Its main advantages include:

◎ Reduced complexity: By integrating multiple functional modules, the Thor chip reduces the number of electronic control units (ECUs) in the vehicle, simplifying system design.

◎ Enhanced reliability: The single-chip architecture eliminates communication delays between systems and improves overall system stability through hardware-level fault tolerance mechanisms.

◎ Improved user experience: The integrated system provides a more consistent user experience, including seamless ADAS functionality and smarter cockpit interaction.

NVIDIA Orin-X to Thor-X-Super Chip Evolution

From Orin-X to Thor-X-Super, NVIDIA’s chips have undergone significant performance improvements to meet the growing demands of autonomous driving.

Orin-X, as the starting point, features 12 CPU cores based on the ARM Cortex-A78AE architecture and an Ampere architecture GPU with 5.2 TFLOPS FP32 performance, making it suitable for basic ADAS (Advanced Driver Assistance Systems) tasks in low-power scenarios. It supports LPDDR5 memory with a bandwidth of 205 GB/s and consumes only 50 watts, making it ideal for applications with strict requirements for real-time performance and safety.

As demand grows, Thor-X brings significant improvements. It features 14 more powerful ARM Neoverse V2 architecture CPU cores, delivering 630 KDMIPS performance, along with a Blackwell architecture GPU offering 9.2 TFLOPS FP32 performance, significantly enhancing multi-tasking capabilities and deep learning inference performance.

Thor-X’s memory bandwidth has been increased to 273GB/s, supporting the LPDDR5X standard, while its power consumption ranges from 70 to 140 watts, balancing high performance with power management. It also supports PCIe 5.0 interfaces, offering higher data transfer rates and more connection options, making it suitable for complex multi-sensor fusion and environmental perception tasks.

The top-of-the-line Thor-X-Super represents the pinnacle of this series. It not only doubles the number of CPU cores to 28 but also further enhances GPU performance to 18.4 TFLOPS FP32 computing power, specifically designed for the most complex and demanding fully autonomous driving systems. Its memory bandwidth reaches an impressive 546 GB/s, utilizing 512-bit-wide LPDDR5X memory to ensure extremely high data throughput.

The Thor-X-Super’s power consumption ranges from 140 to 280 watts, though it imposes higher demands on thermal management, it is capable of handling high-performance computing tasks. Additionally, it supports the latest PCIe 5.0 interface to meet the demands of large-scale data interaction.

From Orin-X to Thor-X-Super, these chips have progressively increased core count, GPU computing power, memory bandwidth, and optimized power consumption and interface expansion capabilities to accommodate a wide range of complex tasks, from basic driver assistance to fully autonomous driving.

● The choice of chip depends on the specific application requirements:

◎ If high performance and power efficiency are not a priority, Thor-X-Super is the best choice;

◎ If cost-effectiveness or power optimization is a priority, Orin-X or Thor-X can be selected.

This evolutionary path demonstrates NVIDIA’s continuous innovation and technological leadership in the automotive chip sector.

Summary

From a technical perspective, NVIDIA’s Thor chip, with its innovative architecture design, exceptional computing performance, and system integration capabilities, sets a new standard for the electric vehicle industry. It not only meets the current demands of intelligent driving but also lays the foundation for the realization of advanced autonomous driving technologies in the future.