Nvidia's "epic" autonomous driving chip is unveiled! Computing power 2000TOPS, compatible with cockpit entertainment functions

Still the same leather jacket, still the familiar old Huang.

At 11:30 last night, the curtain of NVIDIA's Fall GTC opened on time, and NVIDIA founder and CEO Huang Renxun delivered a speech via live online.

For those who are interested in autonomous driving, the biggest highlight of this conference is that NVIDIA replaced the Atlan chip released a year ago with the Thor chip with a computing power of up to 2000TOPS as the successor to Orin.

Of course, as NVIDIA's latest masterpiece in the field of smart cars, the innovative essence of Thor is not simply to increase computing power, but to promote the integration of distributed electronic and electrical architecture of automobiles to a certain extent, moving towards a centrally integrated computing platform.

Interestingly, the Thor chip is expected to be launched globally in 2025 by Geely's high-end electric vehicle brand Zeekr.

Zeekr 001 is the world's first car model to launch Mobileye EyeQ5. Not long ago, there were rumors that "Zeekr will terminate its cooperation with Mobileye."

DRIVE Thor is launched, one chip can be used for multiple purposes

If we make a performance ranking for the autonomous driving chips currently on the market, NVIDIA's Orin chip will definitely be a strong contender for the championship.

In fact, NVIDIA launched Atlan, an autonomous driving chip with a computing power of 1000 TOS, at GTC 2021 in the spring, hoping that it would succeed Orin. Although the official did not explain in detail the reason for replacing Atlan, it is no longer important. The new DRIVE SoC roadmap has already explained everything - Thor will become the successor product of Orin.

However, NVIDIA has announced that Atlan will produce samples in 2023, but did not disclose the specific timetable for the Thor chip.

In the past two years, many new and old car manufacturers have vigorously promoted the Orin chip as one of the selling points of their new models when launching new models. After this press conference, the same situation may soon happen again - but with the Thor chip as the target.

NVIDIA's vision for the Thor chip: Replace the separate computers in today's cars with one computer that can do it all, leveraging functional safety design techniques with extensive isolation (including MIG) to prevent separate tasks from interfering with each other.

Simply put, in addition to powerful computing power, the Thor chip also unifies traditional automotive distributed functions, integrating assisted driving, parking functions, cockpit entertainment, etc., greatly improving development efficiency and the speed of software updates and iterations.

This means that this autonomous driving chip not only integrates driving and parking, but also does almost all the work of a smart cockpit chip.

According to Huang Renxun, a system based on the Thor chip can concentrate all the functions of a smart car on a single AI computer, and "centralization of on-board computing resources can reduce costs by hundreds of dollars."

One chip for multiple uses, multi-domain computing

For car companies, there are multiple ways to configure the computing power of the Thor superchip. They can use all 2000TOPS for autonomous driving functions, use part of it for in-vehicle AI and entertainment functions, and use the other part for assisted driving.

Such technological achievements stem from the multi-domain computing capabilities of the Thor chip.

Multi-domain computing can divide functions such as autonomous driving and infotainment into different task intervals, allowing them to run simultaneously without interfering with each other.

These different types of functions are usually controlled by dozens of electronic control units distributed throughout the vehicle. The emergence of the Thor chip has enabled the integration of vehicle functions, which not only reduces costs, but also solves the problem of lengthy wiring harnesses caused by intelligence, reduces vehicle weight, and says goodbye to distributed ECUs.

The Thor chip can also connect all the vehicle's displays, sensors, etc. to a single chip, greatly simplifying the complex process of automobile manufacturing. Of course, this is also the general trend of the evolution of automotive electronic and electrical architecture from distributed to centralized.

DRIVE Sim simulation platform

In this speech, NVIDIA's DRIVE Sim simulation platform was also its proud product.

According to Huang Renxun, NVIDIA can build 3D scenes based on data recorded by sensors, import these scenes into DRIVE Sim, and enhance them with manually created or AI-generated content, thereby converting relevant data into a simulation environment for closed-loop testing.

Another important feature of DRIVE Sim is hardware-in-the-loop, which means that the in-vehicle software can be run in the AI ​​factory without the AI ​​computer being aware that it is in a test environment.

DRIVE Sim, which supports hardware-in-the-loop, can also simulate the in-vehicle environment. Automotive designers and software engineers can collaborate on development in DRIVE Sim while running all actual computers and software.

Therefore, DRIVE Sim is equivalent to a virtual design studio, which is also a key means for NVIDIA to accelerate the development of AI cockpits and in-vehicle entertainment information systems.

NVIDIA also spares no effort in safety systems and processes. Huang Renxun said: "NVIDIA R&D has invested 15,000 man-years to conduct safety assessments on 5 million lines of code. DRIVE Thor is designed to fully comply with ASIL-D operations and ISO26262 safety specifications."

Thor chip, lightning fast

In terms of speed, the Thor chip achieves a leap forward in accuracy for deep neural networks.

The Thor chip has 8-bit floating point (FP8) capability. With the precision of 8-bit floating point (FP8), the Thor chip introduces a new way of data processing for automobiles.

Traditionally, developers of AV (autonomous vehicle) simulation platforms have needed to move from 32-bit floating point to 8-bit integer data formats, and FP8 precision simplifies this transition, allowing developers to transfer data without sacrificing accuracy.

Chip interconnect technology NVLink‑C2C

At the conference, NVIDIA also launched NVLink‑C2C technology, which can connect two Thor chips at the same time, making the two chips a monolithic platform with a single operating system. The advantage of NVLink‑C2C lies in sharing and scheduling, which can distribute work across links with minimal computing power cost.

Under the development trend of software-defined cars, NVLink‑C2C technology provides cars with sufficient computing power redundancy to ensure the flexibility of software configuration and provides computing power guarantees for future OTA of smart cars.

NVIDIA has great ambitions to be compatible with smart cockpits

Two years ago, the smart cockpit chip market was still in a fierce competition among many players, and car companies placed their bets on different targets to determine who would win in the end.

Until the emergence of Qualcomm's 8155 chip, the smart cockpit chip industry finally welcomed a leader, and car companies gradually regarded the 8155 chip as a symbol of their flagship.

Not long ago, Zeekr spared no expense in upgrading to 8155 for free, which made many friendly competitors grit their teeth and call it "unruly"; it also made consumers exaggeratedly present a banner with the words "conscientious car company, create a model together".

However, NVIDIA's lateral change of lanes - the Thor chip is compatible with in-vehicle entertainment functions - has added a bit of excitement to the smart cockpit chip track.

It is understood that NVIDIA's DRIVE Concierge is the starting point for many functions of the in-vehicle infotainment system (IVI system). It is an AI voice assistant similar to Siri, and its comprehensive capabilities are very powerful.

DRIVE Concierge provides personalized entertainment for every passenger in the car through interactive forms, including graphic visualization in the smart cockpit, intelligent digital assistant, driver and occupant monitoring, and streaming content such as games and movies.

According to Huang Renxun, through a centralized architecture, DRIVE Concierge can seamlessly coordinate driver information, cockpit and entertainment functions, and support the Android Automotive operating system, so automakers can selectively customize and expand their in-vehicle entertainment functions.

A detailed view of the vehicle’s perception and planning is also critical for drivers, and DRIVE Concierge can provide high-quality, 360-degree, low-latency 4D visualization views, allowing drivers and passengers to see the role of AI through 3D graphics.

In terms of specific entertainment functions, DRIVE Concierge allows passengers to watch videos and experience high-performance games anytime, anywhere. Users can freely choose according to their preferences and stream videos or games on any screen in the vehicle.

Additionally, using the NVIDIA GeForce NOW cloud gaming service, passengers can access more than 1,400 games without having to download them.

In addition to entertainment and interactive functions, the customized smart interior also provides additional safety for the vehicle.

DRIVE Concierge uses internal sensors and a dedicated deep neural network to monitor the driver to ensure that they are paying attention while the driver is in control. The feature can also monitor in real time whether passengers are safe and whether they have left valuables behind.

Conclusion

In recent years, NVIDIA has continued to innovate in the field of autonomous driving and has maintained a huge advantage in hardware. The emergence of the Thor chip has raised the computing power ceiling of autonomous driving chips.

The horizontal compatibility of smart cockpit functions also allows the industry to see that the field of smart cockpits will usher in a new possibility.