Towards the Thousand-T Computing Power Era, the Most Powerful Intelligent Driving Chip Thor is on the eve of mass production

At the GTC conference in the fall of 2022, NVIDIA released the next-generation in-vehicle computing platform "DRIVE Thor", which can provide up to 2000 trillion floating-point operations per second, or what we often call "AI computing power", up to 2000 Tops.

Thor's goal is to significantly accelerate the progress of smart cars and autonomous driving and become the central computer of the car.

Compared with the previous generation OrinX, it has an 8-fold performance improvement, which also shows NVIDIA's determination to "squeeze the toothpaste and accelerate everything."

After the launch of Thor, the discussion about whether autonomous driving really requires such "great effort" lasted for quite a long time.

This situation did not change significantly until after the 2024 Beijing Auto Show. Some Tier 1s received intensive inquiries from OEMs regarding the development of Thor after the auto show, and NVIDIA also officially announced its cooperation with many customers at this auto show.

This situation mainly comes from two reasons:

• 
  

• In the first half of 2024, the "end-to-end" data-driven development approach will become a consensus in the intelligent driving industry, and OEMs will seek platforms with greater computing power to promote the use of large models in vehicles.

• 
  

• Leading OEMs hope to take advantage of platforms like Thor to gain an advantage, while traditional OEMs such as BYD hope to catch up or surpass them as quickly as possible, which perfectly constructs the logic of an arms race.

According to incomplete statistics from HiEV, so far, OEMs such as Ideal, Xpeng, Zeekr, BYD, and Jiyue have reached cooperation with NVIDIA and are developing intelligent driving computing platforms based on Thor; among suppliers, Lenovo Automotive Computing, Desay SV, and Zhuoyue Technology are also developing domain controllers or intelligent driving solutions based on Thor.

On the eve of mass production of the most powerful chip for intelligent driving, we talked with major Tier 1s such as Lenovo Automotive Computing and Desay, as well as upstream and downstream companies in intelligent driving:

• 
  

• The 700T Thor is becoming the model that passenger car OEMs are most interested in.

• 
  

• The earliest mass-produced model equipped with Thor may be released in the first quarter of next year;

• 
  

• Even at the 1,000-T computing power level, Thor is still not a complete cabin-and-pilot integrated computing platform;

• 
  

• Thor X may have a significant impact on the commercialization of Robotaxi.

1. End-to-end + big model: the anxiety and hope of OEMs

Big models are sweeping smart driving and smart cockpits. This new development method has accelerated the consumption of AI computing power and greatly stimulated the anxiety of OEMs about the process of intelligence.

How much computing power does a large model consume?

In 2021, a new brand OEM was the first to use Qualcomm 8295 for cockpit development. At that time, the computing power of the 8295 with a single AI core was already 30 Tops.

Due to the abundant supply of computing power, the OEM plans to develop a single-lane driving function on the 8295 in addition to the cockpit function as a redundant backup for the intelligent driving system.

However, by the second half of 2022, the computing power of 8295 was already tight due to the development of the core voice model based on a single AI core. The OEM finally obtained the license for dual AI cores from Qualcomm, increasing the AI ​​computing power of 8295 to 60 Tops, but the mass production of redundant intelligent driving functions was shelved.

The same will be true for intelligent driving computing platforms in 2023.

The dual OrinX configuration used by most domestic OEMs in high-end solutions refers to the design of Tesla's HW3.0 dual FSD chips to a certain extent, with one chip initially used as a redundant backup. But in the second half, the computing power of the dual OrinX became very tight.

Take Ideal Auto as an example. Ideal Auto started the internal testing of the end-to-end intelligent driving solution in OTA 6.1.0, which adopted a dual-system design of "System 1 + System 2". System 2 is mainly a VLM visual language model with a parameter scale of 2.2 billion (2.2B), and the frequency of VLM is about 3-4 Hz.

Just for the deployment of the VLM visual language model on OrinX, Ideal has spent a lot of effort on model compression and operation optimization, and this is only the initial stage of installing the intelligent driving large model on the car.

An AI practitioner told us that the parameter scale of a truly large model is above 10B. The intelligent driving system has high real-time requirements, and the system operating frequency must be guaranteed to be at least 10-20 Hz.

Smart driving and smart cockpits have reached a consensus on the development methods around large models, which all point to a stronger next-generation computing platform optimized for large AI models.

In the context of the extreme scarcity of high-powered in-vehicle computing platforms, Thor has become the "hope of the whole village" for large-scale models on vehicles.

2. From L2+ to L4, Thor’s ambition

Although Thor was released in 2022, there is still little public information about this powerful computing chip. This is closely related to the development rules of the automotive industry.

In 2022, NVIDIA's first Thor launched had a computing power of 2000 Tops.

However, as its development progressed and driven by demand from major customers, Thor gradually evolved into 4 main versions:

• 
  

• Thor X, 1000T computing power;

• 
  

• Thor S, 700T computing power;

• 
  

• Thor U, 500T computing power;

• 
  

• Thor Z, 300T computing power.

Among them, the Thor S with 700T computing power has become the configuration that passenger car OEMs pay the most attention to.

Unlike the common perception that Thor is very expensive, the computing power of a single chip is greater than that of dual OrinX, and the inter-chip communication problem is reduced, which reduces the system complexity. Several people close to NVIDIA revealed to HiEV that the domain control system cost of Thor S is lower than that of dual OrinX, and the performance is better.

As for the 1000T Thor X, the main demand currently comes from L4 autonomous driving companies.

Whether it is Waymo, Cruise, or China's Pony.ai and WeRide, they are all seeking to expand the operation of Robotaxi fleets in larger areas.

The core issue that needs to be solved for large-scale fleet operations is how to reduce the cost per vehicle and ensure the stability of the long-term operating system.

Thor X is the automotive-grade, high-computing computing platform on the market that is most in line with this demand.

Lenovo Automotive Computing recently launched the domain controller AD1 based on Thor X, which is developed for the pre-installed mass production of Robotaxi, including WeRide. AD1 contains two Thor X chips and has a computing power of up to 2000T.

In addition to Robotaxi companies, a small number of car companies targeting the development of L3 intelligent driving systems are also considering using Thor X.

With thousands of terabytes of computing power, in addition to intelligent driving, it is also possible to consider running some large language models for smart cockpits.

Tang Xinyue, vice president of Lenovo Group and head of Lenovo's automotive computing business, told HiEV:

“We believe that the 700T version (Thor S) will be the first to be used. It can meet the requirements of high-end models above 300,000 yuan for intelligent driving and cockpit experience, and can perform cross-domain computing.

The 500T (Thor U) and 300T (Thor Z) versions are aimed at the needs of low-end market models, and have greatly improved ADAS performance (compared to existing solutions), which is incomparable to the previous Orin versions.

Even the lowest version of Thor, 300T is better than dual OrinX. "

In NVIDIA's intelligent driving computing product series, compared to Orin, which mainly focuses on the high-end L2+ market for passenger cars, Thor has greater ambitions and more comprehensive coverage.

Whether it is the future large-scale deployment of L4, the experience upgrade of the existing urban NOA and high-speed NOA systems, or the upcoming L3 mass production, they are all within its range.

But one thing that is different from NVIDIA's expectations is that Thor is not the central computer of a smart car as originally planned.

With the arrival of "end-to-end + big model", people have found that the growth rate of the upper limit of intelligent driving systems is still rapid.

When the intelligent driving problem itself has not been well solved, 1000T will not be the end of automotive computing power, at least for the highest-end, flagship models.

3. Lenovo and Desay, arms dealers in the era of big models

Tang Xinyue, head of Lenovo's automotive computing, was responsible for automotive computing-related businesses at Intel and Flextronics in his early years.

Around 2017-2018, his business line provided on-board computers to top Robotaxi companies at the time, such as Cruise and Argo.

In 2022, when Lenovo formed a team and entered the business of automotive computing products, Lenovo quickly decided to target the DRIVE Thor generation of products and signed a license with NVIDIA.

In August this year, Lenovo Automotive Computing's automotive-grade domain controller product AD1 based on Thor X rolled off the production line at the Hefei Industrial Base. This means that next year, Robotaxi, which has been explored by the industry for nearly 10 years, will enter the initial stage of large-scale commercial operation.