NVIDIA releases dominant chip Thor, with computing power of 2000TOPS! !
Latest update time：2022-09-21
Reads：
The Thor SoC is here.
Author |
Juice
Edit |
Xiaohan
Nvidia throws another "nuclear bomb", the new smart car chip Thor is officially released!
This SoC chip has 77 billion transistors inside, which can achieve an AI computing power of 2000TOPS, or 2000TFLOP.
▲Basic performance of NVIDIA Thor chip
Because the parameters were too powerful, Nvidia simply named it Thor. Yes, that’s the man who swings the hammer in Marvel.
Because the parameters are too powerful, Thor SoC has even directly replaced the new product originally planned for mass production in 2024/2025, which is the current next-generation product of Orin, Altan SoC (1000TOPS).
Because the parameters are too powerful, Nvidia boss Huang Renxun did not even describe it as an "autonomous driving chip". Instead, he made it clear that this SoC was born for the central computing architecture of the car. Use this chip to build a controller that can simultaneously Provides computing power for multiple systems such as automatic parking, smart driving, vehicle control, instrument panels, and driver monitoring.
The meaning is very clear, one is worth six.
▲NVIDIA Thor supports cockpit integration
Lao Huang said that Thor SoC will be mass-produced in 2024. And Geely’s electric vehicle brand Ji Krypton was the first to announce that it will equip its models with Thor SoC starting from 2025.
With the rapid development of automobile intelligent driving technology, NVIDIA has a very strong position in the mid-to-high-end autonomous driving chip market in the past two years. Mid-to-high-end models such as Weilai ET7, Ideal L9, Xpeng G9, Zhiji L7 and other mid-to-high-end models have basically chosen NVIDIA. Orin autonomous driving chips are used to develop high-speed and urban NOA automatic navigation assisted driving systems.
While Orin is gradually seizing the market this year and next, NVIDIA is also actively developing next-generation products to maintain and expand its market position in the smart car field. Last spring, it released Atlan, whose AI computing power of 1,000TOPS per chip was already astonishing.
Unexpectedly, Nvidia wanted to play a big game again. Atlan was directly replaced by Thor, and it produced a "nuclear bomb" level product with a single 2000TOPS. It is simply invincible at present.
And it is worth noting that Raytheon no longer focuses on the old brand of autonomous driving chips this time, but uses its super computing power to directly set the benchmark for central computing main chips. When the EE architecture for central computing actually begins mass production, NVIDIA will once again gain an overwhelming advantage by virtue of its first-mover advantage, and at the same time, it will also kill the market for other small computing power chips.
It has to be said that Nvidia is not only at the forefront, but also the first to move to a new track.
Benefit of this article:
Nvidia throws a nuclear bomb, and the self-driving chip industry accelerates its involution. Share the report
"Smart driving chips are emerging, and autonomous driving is in the ascendant"
. Reply in the dialog box
[
Che Dongxi 0411
]
to download the report.
01
.
Released new Thor chip to support smart driving and smart cabins
Thor can support both ADAS systems and IVI systems, has 77 billion transistors, and has a computing power of more than 2,000 TOPS.
Huang Renxun introduced that there are three main points to achieve this goal, namely the upgrade of the CPU (Grace), GPU (Ada Lovelace) and the engine (Hopper) that processes the Transformer model. Hopper provides the amazing Transformer engine and Vision Transformer's rapid transformation, while Ada is Nvidia's latest GPU product, built on the 4nm process. Its multi-instance GPU invention will help centralize on-board computing resources and reduce costs. Down hundreds of dollars.
▲Nvidia CEO Jensen Huang
Grace CPU was also introduced in Thor, which also has good performance. In the past, all parallel algorithms were accelerated by NVIDIA GPUs. Other workloads were often limited by single threads, and Grace happened to have very good single-thread performance.
Based on the above foundation, NVIDIA engineers built Thor, which also contains many innovative designs.
Thor can be configured in multiple modes, and all of its 2000TOPS and 2000FLOPs can be used for autonomous driving workflows; its 2000TOPS computing power can also be used separately, such as one part for cockpit AI and infotainment systems, and the other part for autonomous driving. .
▲The evolution process of Nvidia’s autonomous driving chips
Thor's multi-compute domain isolation allows concurrent, time-sensitive multi-process, non-disruptive operation to run Linux, QNX and Android simultaneously on a single computer.
This product concentrates many computing resources, reduces cost and power consumption, and achieves a leap in functionality.
Currently, a car's parking, active safety systems, driver monitoring, camera mirroring, clustering and infotainment are all controlled by different computing devices. In the future, these functions will no longer be controlled by separate computers, but will run simultaneously on Thor, and the functions provided by the software will continue to be improved over time.
▲NVIDIA Thor supports cockpit integration
The advantages brought by this product are also very obvious. It can significantly simplify the automotive EE architecture and alleviate supply constraints. On the other hand, it can reduce the wiring scale of the product and reduce the weight of the vehicle, thereby reducing costs.
Jikrypton Motors will begin to equip Thor for its next-generation vehicle model platform in 2025, which will continuously improve the performance and experience of the Jikrypton fleet, and its safety and intelligence capabilities will also be continuously upgraded.
▲NVIDIA Thor will be the first to land on JK Automobile
Jikrypton's adoption of Thor also proves the recognition of automobile manufacturers for cost-effective central computing architecture. This architecture can simultaneously meet the needs of safe, reliable and highly automated driving capabilities, and can also meet the complex and heavy workload behind the increasingly rich functions of in-vehicle infotainment systems. Calculation requirements.
02
.
Simulation capabilities have been greatly improved, allowing car companies to conduct virtual development
Jen-Hsun Huang also introduced NVIDIA DRIVE, an end-to-end platform for autonomous vehicle development and deployment created by NVIDIA. On the development side, DRIVE includes Reolicator synthetic data generation, NVIDIA AI infrastructure such as DRIVE Sim and DRIVE Map.
In terms of deployment, NVIDIA DRIVE includes full-stack driving and in-vehicle AI applications, AI computers and the Hyperion autonomous vehicle reference architecture.
NVIDIA staff developed an AI workflow that can build 3D scenes based on recorded sensor data. After the 3D scenes are imported into DRIVE Sim, they can be enhanced with human-created content or AI-generated content.
This video-to-3D geometry workflow runs on NVIDIA OVX systems, allowing NVIDIA to create simulation scenes on a global scale.
For example, NVIDIA uses the Neural Recinstruction Engine to provide support for DRIVE Sim. In a few minutes, the engine can reconstruct a complete 3D digital twin of the driving record based on sensor data.
▲NVIDIA DRIVE Sim can create 3D digital twins in real time
Through AI technology, NVIDIA can collect and reconstruct objects. These materials will be loaded into Omniverse (a computer graphics and simulation platform previously released by NVIDIA) and can be used in DRIVE Sim at any time, with the assistance of DRIVE Map. Next, developers can place dynamic objects such as vehicles or pedestrians, and then the simulation system can change the environment and conduct closed-loop testing to avoid more risk scenarios.
Even developers can create new scenes based on the collected scenes and materials, and can generate synthetic ground truth data to train the perception network. These scenes can be used for end-to-end testing.
▲NVIDIA DRIVE Sim can create extreme scenes by yourself
Huang Renxun said that DRIVE Sim plays a big role in building autonomous driving systems and is a very important part of the CI/CD continuous integration/continuous deployment process.
DRIVE Sim also supports in-car environment simulation. Future cars will not only have simple dashboards, but also surround displays for digital and physical design. Automotive designers, software engineers and electronic engineers can collaborate in DRIVE Sim while Run all the actual computers and software stacks.
▲Nvidia’s concept for the car cabin
DRIVE Sim will become a virtual design studio for these engineers.
In the development process of robots and autonomous driving, safety is very important. NVIDIA has invested 15,000 man-years in safety systems and processes and conducted safety assessments on 5 million manual codes. The design of its autonomous driving chips and platforms all follow Establish common safety standards in the industry.
Nvidia has made great progress in developing end-to-end autonomous driving. As can be seen in the video displayed at the scene, after the user arrives in the car, the AI assistant will report today's itinerary to the user, and then complete a series of driving and parking actions by itself. No human participation is required in the entire process.
▲NVIDIA’s autonomous driving system can achieve end-to-end
At present, NVIDIA's self-driving chip products have been launched in more than 40 car companies, car companies, trucks, autonomous taxis, and autonomous minibus companies. Among them, Xpeng Motors' latest flagship model G9 uses Orin chips. This The car will be officially delivered later this year.
After adopting the Orin chip, this car will have highly advanced assisted driving functions, such as autonomous driving and autonomous parking on major and secondary city streets, highways and private roads, and the ability to automatically enter parking lots. , special scenes such as narrow streets and toll stations in the city.
03
.
Conclusion: Nvidia expands the boundaries of its automotive business
At previous GTC conferences, NVIDIA's focus in automobiles has always been around autonomous driving, and it has regarded autonomous driving as a very important market growth point.
At this conference, although autonomous driving is still the highlight of Nvidia, it has also increased its layout in smart cockpits, simulation systems, cloud and other aspects. It has changed from wanting to help car companies do autonomous driving to wanting to Help car companies build good cars.
At present, whether it is Tesla, overseas established chip factories, or domestic technology companies and startups, their product capabilities have significantly improved in recent years, and they are also trying to shake Nvidia's position. Although in the short term, they will not have any impact on Nvidia's position. Nvidia has no impact, but it is still necessary to be prepared for danger in times of peace.
As NVIDIA gradually increases its presence in the automotive business, its position may become more stable.
Scan the QR code below to watch the video replay! !
京公网安备 11010802033920号