NVIDIA Launches NVIDIA Thor, a Centralized Vehicle Computer with Revolutionary Performance
Author: Tian Hui
Editor: Zhu Shiyun
In the automotive industry, the only one who can surpass NVIDIA is NVIDIA itself.
At the GTC conference on September 20th, NVIDIA founder Huang Renxun announced the launch of NVIDIA Thor, a centralized vehicle computer with a computing power of 2000 TOPS+2000 TFLOPS which can satisfy the needs of almost all vehicle computing requirements for features such as autonomous driving and intelligent cabins.
Thor’s computing power is eight times that of existing Orin chips, and 14 times that of Tesla FSD chips.
If 2000 TOPS is still not enough, Thor also supports NVLINK C-C function, which can double performance.
At the spring GTC conference in 2021, the Atlan chip with 1000 TOPS computing power was abandoned by NVIDIA before it was delivered and directly replaced by the Thor chip. It is expected to achieve mass production in 2025, and perhaps a Chinese brand Extreme car-maker will be the first to adopt it.
In addition, NVIDIA also showcased a new feature of NVIDIA Sim, an AI workflow named Neural Reconstruction Engine, which can assist car manufacturers in building simulation environments quickly and improve autonomous driving simulation training capability.
Through software and hardware innovation, NVIDIA is quietly transforming itself from a supplier of automotive chips to a core player in the future automotive industry.
What does Thor mean for the era of autonomous driving? The real question is, will NVIDIA’s Thor sell well?
Thor is the V12 Engine of the Autonomous Driving Era
Thor’s Chinese name means the God of Thunder, which shows how powerful it is.
In terms of performance, Thor SOC has 77 billion transistors which can achieve a computing power of 2000 TOPS for AI or 2000 TFLOPS for floating-point operations.
For comparison, the computing power of the Xavier chip used by the current XPeng P7 is 30 TOPS, and that of the Orin chip used by the NIO ET7 is 254 TOPS. Thor’s 2000 TOPS computing power is undoubtedly a revolutionary leap in performance.
For those in the automotive industry who are unfamiliar with units of computing power such as TOPS, let’s make a comparison using automotive industry terms: Xavier-30 TOPS, Orin-254 TOPS, and Thor-2000 TOPS.Xavier = three-cylinder engine, Orin = four-cylinder engine, Thor = V12 engine.
To create the autonomous driving “V12 engine” Thor, Nvidia has undergone a significant upgrade to the Thor chip architecture.
Within the Thor chip, there are three advanced chip architectures from Nvidia: the Grace GPU, the Hopper (Transformer engine), and the Ada Lovelace GPU.
The Hopper (Transformer engine), which is like the 8AT automatic transmission of a car, is a newly added and vital architecture.
During the speech, Huang Renxun mentioned that the amazing Transformer engine and Vision Transformer’s rapid transformation of the Hopper architecture are essential and must be included in the next generation of robot processors.
The advantage of the Transformer engine is that it can convert FP8, FP16, and FP32 mixed-precision formats to FP8 precision format without losing accuracy, improving computational efficiency.
The Transformer engine’s role in AI computing is similar to that of the 8AT in cars, both for more efficient operation. Nvidia’s official data shows that Thor can improve the inference performance of Transformer deep neural networks by up to nine times.
Grace, on the other hand, is an ARM-based CPU architecture originally developed by Nvidia for servers, and it has now been ported to the centralized vehicle computer Thor. Grace has excellent single-threaded performance and a leading edge when it comes to processing large amounts of data.
Ada Lovelace is the GPU architecture used in NVIDIA’s new RTX4090 graphics card. According to Huang Renxun, at the launch event, the RTX4090 with Ada Lovelace architecture delivers twice the performance of the previous generation graphics card under the same power consumption.
As a result, Grace, who excels at logical judgment, supports top-level decision-making and instruction distribution for intelligent cars;
Ada Lovelace, who excels at rendering, supports various graphics display requirements for intelligent cockpits;
Hopper, designed for AI-computing with heterogeneous data, focuses on autonomous driving requirements.
With 2000 TOPS of computing power, Thor outperforms autonomous driving chips and becomes the central computing chip that can meet nearly all computing power requirements for cars.
Thor provides the foundation for centralized computing
Thor makes it possible to implement centralized electronic and electrical architecture for computing.
Currently, the latest generation of intelligent cars mostly adopt domain-controlled electronic and electrical architectures and, based on the middleware and operating system software capabilities, have achieved cross-domain fusion for some functions.
Although domain control has significantly reduced the control units and shortened the communication bundles, and greatly improved the signal transmission speed compared with the era of distributed architecture with ECU in the past, domain control still has limitations in further functional fusion and improving real-time response as the ultimate goal of intelligent cars with computers on four wheels.
Thor provides the technical foundation for centralized computing.
One fundamental reason why Thor is called a centralized car computer is that it has real-time computing power distribution ability, which can simultaneously apply computing power to functions like autonomous driving, intelligent cockpits, automatic parking, active safety, and so on.
At the launch event, Huang Renxun mentioned that car parking, active safety, driver monitoring, camera mirroring, clustering, and entertainment will no longer be controlled by separate computing devices but by software running on Thor, which can be updated over the air (OTA).
Thor can be configured for multiple modes, including using its 2000 TOPS and 2000 TFLOPs entirely for autonomous driving workflows, or allocating a part for AI in the cockpit and entertainment and part for assisted driving.“Thor’s multi-computing domain isolation allows for concurrent, time-sensitive, uninterrupted operation of multiple processes, including Linux, QNX, and Android on a single computer,” said Huang Renxun at the release event. Thor consolidates numerous computing resources, not only reducing costs and power consumption, but also achieving a leap in functionality.
Software-defined vehicles are the inevitable direction of the automotive industry, and Thor provides the foundation for software updates. In a distributed architecture, the development of software upgrades for different functions requires research and development of different subsystems. In a centralized computing architecture, functionality upgrades can be developed on Thor alone.
The actual benefit Thor can bring to consumers is faster software iteration speed and more frequent functionality upgrades.
From the perspective of the automotive industry, a high-performance chip that solves all of the computational needs of an automobile is the direction of the industry. However, few automobile manufacturers have put it into practice in the automotive industry, and most automobile manufacturers are still on a distributed architecture. The reason behind this is that there isn’t currently a chip that can truly achieve multi-domain integration.
Chinese driver assistance domain controller manufacturer Desay SV Automotive launched the ICP Aurora system in the first half of this year, which integrates core functions such as intelligent cabin, autonomous driving, and networked services, enabling the leap from “domain control” to “centralized computing.”
However, publicly available information shows that ICP Aurora is a system-level multi-chip integration solution that combines current mainstream high-performance chips such as Orin, SA8295, and A1000. It is less efficient than Thor’s chip-level fusion solution.
But if this system can be equipped with Thor chips in the future, it will greatly optimize the system structure and make it easier to mass-produce.
Will NVIDIA’s Thor still sell well?
At the 2022 Spring GTC Conference, NVIDIA announced its Orin chip collaboration with BYD.
At the 2022 Fall GTC Conference, NVIDIA announced that Extreme Vision would be the first automaker to use Thor.
Achievements are achievements, and business is business. Will Chinese automakers support NVIDIA again this time, as they did with Orin?
“For the Chinese automotive industry, Orin has only just begun to be mass-produced in vehicles. NVIDIA’s announcement at this time that it’s abandoning the development of next-generation Atlan chips in favor of the more aggressive development direction of centralized vehicle computing systems like Thor is somewhat unexpected.”
After the press conference, Huang Renxun explained during the global media communication meeting that:
I decided that we would delay for a new chip by a couple of months and bring in all of the amazing new technology.
So that we can have something amazing that takes advantage of all these new technologies that are in Hopper, in Grace, in Ada as sooner to the market.
I just couldn’t imagine waiting another two more years.
Huang Renxun’s answer was straightforward: abandoning Atlan for the release of Thor would only require a few months of delay, whereas postponing Atlan and releasing Thor simultaneously would mean a two-year delay. The reason for this decision is that the new technology will bring amazing things.
In terms of computing power, Thor has the ability to change the electronic and electrical architecture of automobiles and make astonishing changes. However, in the Chinese market, there are more choices among automobile companies, and whether they will follow NVIDIA’s chip iteration rules is still unknown.
Apart from NVIDIA, Chinese automobile companies can also choose Horizon, Qualcomm, and other self-driving chip platforms that combine high computing power and openness.
Horizon J5 has signed cooperation agreements with multiple automobile companies, and the upcoming Ideal L8 is expected to be the first car to be launched on Horizon J5.
Qualcomm Snapdragon Ride has announced a cooperation agreement with HoloMatic, a subsidiary of Great Wall Motors, and both parties will conduct advanced assisted driving research on this platform.
Apart from these, NVIDIA’s old rivals Mobileye and Texas Instruments also have assisted driving calculation chips.The advantage of NVIDIA lies in its computing power as well as its development platform and toolchain. By using NVIDIA’s complete set of toolchains, automakers can save time and cost, and simplify development workload in autonomous driving research. However, automakers also need to endure the impact of NVIDIA’s high power consumption on the car’s cruising range.
After earning the trust of automakers with Orin chips and entering their supply chains, NVIDIA launched the Thor chip, which provides even higher computing power to meet almost all the computing requirements of cars at once. It is somewhat risky.
Currently, large-capacity chips represented by Orin have only recently been used in cars, and all automakers have not yet fully developed high-speed and city-level navigation assistance functions nor have they encountered computing bottlenecks. Even in the coming years, Orin’s computing power will be able to meet the needs of automakers.
Under these circumstances, automakers may not be eager to switch to Thor chips.
In addition, the application of Thor chips means a fundamental change in the vehicle’s electronic and electrical architecture, which is a major change for automakers that affects the entire vehicle. The entire vehicle platform adapted to the new architecture needs to be changed, which will be a huge cost. Obviously, it is more appropriate to use Thor when developing brand-new models.
The procurement cost is also important.
Orin’s selling price is around $300, does 2000TOPS Thor imply the price of 8 Orin chips?
Having NIO install 4 Orin chips in the automobile industry has already been seen as an “impressive move” by investors. Even if the cost of the previous Qualcomm 8155/8295 is included in the price of 8 chips, whether automakers can accept them in large quantities remains to be seen.
Also, Chinese automakers seem to have a natural interest in self-sufficiency, so should automakers develop chips themselves? After all, US Tesla did it.
In any case, Nvidia released the centralized on-board computer Thor when Orin was just being mass-produced and installed in cars, which is both expected and reasonable.
However, given the leading position of China’s intelligent electric vehicle industry in the global market and NVIDIA’s leading position in the field of autonomous driving chips, the powerful alliance between China’s leading automakers and NVIDIA is still worth looking forward to.
A powerful alliance means that consumers are a step closer to experiencing autonomous driving.
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.