Follow and star the electric planet News
Punch in every day to read
Gain a deeper understanding of the automotive industry transformation
————————
Producer: Electric Planet News
Author: Yu Fei
Late last night, NVIDIA GTC 2022 kicked off with a concept Mercedes.
Obviously, cars, or smart cars, have become the core topic of this GTC, and have become the core business of NVIDIA in the future.
After being snatched away by Tesla D1 last year with the title of "strongest AI computing", NVIDIA released the H100 chip this year, regaining the AI performance throne, and also released the world's strongest AI training server - EOS, with a computing power of up to 18.4EFLOPS.
In addition, NVIDIA's circle of friends ushered in an "old friend" that we are all familiar with: BYD. Huang Jenxun, founder of NVIDIA, said that BYD's new car equipped with NVIDIA Orin chips will be released in the first half of 2023.
In addition, Nvidia also released a new "public" self-driving hardware Hyperion 9, with 14 cameras, 20 ultrasonic radars, 6 millimeter wave radars, and 3 lidars.
Lao Huang claims that he only needs to buy a set of solutions for his family, and he can immediately have the L4 level of automatic driving capabilities.
Completing NVIDIA's "autonomous driving universe" is the self-developed high-precision map Drive Map, which claims to be about to complete the mapping of 500,000 kilometers of roads in North America, Europe and Asia, and will enter the global market.
We have all heard the myth of Musk's richest man, but in fact, Huang Jenxun and his NVIDIA are also a microcosm of the new stage of the Internet and the new stage of the automotive industry: the third-party suppliers who currently have the most complete product system for assisted driving computing.
As a T1 supplier that doesn't build real cars, Nvidia essentially said only one thing last night:
The infrastructure required for smart cars, I can do it alone.
We also interviewed Danny Shapiro, vice president of NVIDIA's automotive business remotely yesterday, and today we will talk about NVIDIA GTC 2022 in conjunction with what we interviewed.
First, what is the "infrastructure" of smart cars?
Weibo, WeChat, Douyin, these are the three major national social apps.
The reason why they are mentioned in this article is because they, and the apps they represent, constitute most of the "infrastructure" of China's mobile Internet - information circulation, interpersonal socialization, and group agglomeration, all based on this.
The reason why it is said to be "basic" is also because the main entrepreneurial direction of China's mobile Internet is also based on these social apps.
Back to smart cars, the focus of our discussion is tilted from software to hardware, but it can also be compared: the hardware that determines the direction, use, and even functional structure of smart cars can be considered the "infrastructure" of smart cars.
If we define it in more detail, automatic driving sensors, autonomous driving computing platforms, autonomous driving training servers, and self-driving high-precision maps, these are the "infrastructure" needed in the driving field of smart cars.
The focus of our article today is that Nvidia has played new tricks on these "infrastructures" in 2022.
Second, from BYD to Mercedes-Benz are used
At gtcover last fall, NVIDIA brought the first "public" intelligent driving all-inclusive solution in the history of the automotive industry: Hyperion 8.
The reason why I dare to say "the first model of the automobile industry" is because Hyperion 8 is a set of meticulous, afraid that car companies have a little knowledge blind area nanny-level intelligent driving solution:
Don't know what sensor to use? It's okay, 12 cameras, 9 millimeter wave radar and 1 lidar are all equipped for you, and even given a specific model, you don't have to be afraid of me shoddy charging; buy hardware do not know how to use? It's okay, the software and the model are all matched for you, as long as you give money, I will help you adjust.
This is the "nanny level".
At this year's GTC, Huang Jenxun officially confirmed that the final mass production version of Hyperion 8 will be equipped with two Orin computing chips, reaching the computing power of 508TOPS, supporting L3 level intelligent driving, and will be officially installed in 2024 - and the first user of Hyperion 8 is Mercedes-Benz.
While the iron is hot, Nvidia also released Hyperion 9 last night, a nanny-level solution that supports L4 level autonomous driving, and is expected to be loaded in 2026.
Compared to the 8, the dual-chip 9 was upgraded to Atlan, orin's successor, based on TSMC's 5nm process, with unknown parameters. But there are rumors that Atlan's single chip hash rate exceeding 1000 TOPS.
In addition, the Hyperion 9's standard sensors reach a crazy 14 cameras, 20 ultrasonic radars, 9 millimeter-wave radars, and 3 lidars.
How crazy is that? Compared with NVIDIA's solution, tesla currently has only 8 cameras, and LiDAR such as Weilai Volvo is only equipped with 1. But given the performance of a dual Atlan that could exceed four Orins, it's no surprise that sensors fill the body.
What is very special is that even though Nvidia released a one-stop solution, Danny Shapiro said in an interview with us that car companies can decide for themselves what to buy from NVIDIA.
He believes that car companies have a growing demand for intelligent driving, not only hardware, but also a full set of software. Nvidia also wants to expand its business scope and "generate revenue other than hardware".
And for those car companies that want to develop self-driving autonomous driving, NVIDIA still provides pure computing chips, such as Wei Xiaoli, such as LUCID and BYD, which have just entered the game.
According to Huang Jenxun, the first BYD models equipped with Orin chips will be released in the first half of 2023. If you decide to hold on the sidelines because of the recent price increase, you can't help but make a thorough party.
After last night's GTC conference, NVIDIA's car circle of friends is roughly as shown in the following figure, and half of the automotive industry is in the pocket. Interestingly, there are nearly half of the Chinese names here:
Third, NVIDIA brand high-precision map, can it be used?
DRIVE Map, which is the name of NVIDIA's high-definition map. Since last night, DRIVE Map has officially entered the commercial countdown and is currently tentatively scheduled to go live in 2024.
To talk about NVIDIA's high-precision map, we must start with "three camps, two ends, and two engines".
Cameras, millimeter-wave radar, and lidar, which are the three camps of intelligent driving sensors; the car terminal and the cloud, which are the two major computing units that constitute drive map mapping and correction; and the dedicated engine and crowdsourcing engine, which are the two major methods of DRIVE Map mapping.
The first is the three camps, which form the final form of drive map.
According to NVIDIA's official documentation, cameras are used in drive Maps to determine "driving boundaries", such as lane lines, arrows, utility poles, road boundaries, traffic lights, and so on:
Millimeter-wave radar is suitable for poorly lit scenes and suburbs where typical navigation data is less useful:
LiDAR is what NVIDIA believes is the most accurate environmental mapping tool, which can build the world at a resolution of 5 centimeters, "providing the most accurate and reliable representation of the environment."
Ultimately, drive map will build the road traveled by the car into the following picture, so that your car and the server in the cloud can read it, and then apply it to actual driving.
And, as shown in English in the figure above, the NVIDIA DRIVE Map is generated in real time, which is critical.
As mentioned above, DRIVE Map uses two mapping engines – the Deep Map engine for a dedicated mapping vehicle and the crowdsourced map engine for a production vehicle.
The dual engine ensures the efficiency and data volume of map mapping, at the same time, NVIDIA is also equipped with a set of map real-time calculation tools in the production vehicle and server, which can divide the information collected by the sensor to form a complete map at a higher speed.
So the problem of real-time generation and crowdsourcing mapping comes: privacy.
Third-party mapping has always been a sensitive topic for countries, not to mention the mapping of cars running around the road to obtain data and then upload it to the cloud – especially in Europe and China, where privacy policies are detailed.
If you want to enter the Chinese market, how can NVIDIA prove that it is more reliable than Gaode Baidu?
In an interview, Danny Shapiro also did not explicitly disclose which countries Nvidia has been authorized, saying only that NVIDIA "will respect the privacy policies of each country.".
Fourth, the strongest supercalculation on the surface, anti-killing Tesla
The D1 chip on Tesla AI Day last year raised the world's imagination of artificial intelligence to a full level.
When it comes to Tesla, Qualcomm and other chip fields or new or old opponents, Danny said very generously: "Of course, we welcome competition, competition is a good thing for us."
Our interview with Danny was on the morning of the 22nd Beijing time, and gtc. was late at night on the 22nd, and the reason why he was confident was that we knew it more than a dozen hours later: the NVIDIA H100 AI chip.
The H100 shoulders many titles: the first AI chip manufactured by TSMC's 4nm process, the first chip of NVIDIA Hopper architecture... Wait a minute. Its most important task is to help Nvidia regain the performance throne from the Tesla D1.
The H100 did.
Let's compare the D1's BF16 precision hash rate of 362T, while the H100's BF16 precision hash rate is as high as 1000T, and the hash rate in sparse mode can even be doubled to 2000T. What's more, with nearly 6x performance and larger memory, the H100 consumes less than 2x the power of the D1.
However, Nvidia achieved a single-chip counter-kill, and Tesla maintained a leading position in single-cabinet performance with its exquisite structural design.
Last year, Tesla built a "DOJO" server with a D1 chip , the DOJO Pod. The BF16 precision hash rate of a single DOJO Pod is 1EFLOPS, which is scary enough, although it does not achieve the FP32 precision 1E hash rate that Musk said.
This year, Jen-Hsun Huang assembled the H100 chip into a DGX H100 Pod, which also advertises 1E hashrate, but the calculation accuracy is lower than FP8.
However, Huang Jenxun did not concede defeat, and last night he pulled out the big killer: EOS - not the Canon.
EOS is NVIDIA's newest and strongest custom server. Its FP8 accuracy hash rate of 18.4 EFLOPS - 3 times faster than japan's "Fugaku", the world's fastest supercomputer!
For example, Tesla Autopilot is still using Nvidia's old server to run a neural network model (Dojo has not yet built) - Tesla's AI department head Andrej said last year that they looked for Nvidia's customized server, and the computing power is about "only" 1.8EFLOPS.
Changing.
Fifth, OEM and self-research, to what is the boundary?
After talking about a bunch of numbers and terms, we conclude with a brief summary.
Luo Yonghao once complained: "What X are the OEM installing", scolding the mobile phone factory.
Is that correct? By the way, Google, Qualcomm, Samsung, sony these four almost all the core software and hardware of the phone package, so Luo Yonghao has the right place.
However, the deep customization of operating systems, hardware between the driver, tuning, color and signal coordination, and even quality control, this is also a unique effort of major mobile phone manufacturers.
If Luo Yonghao really had a good word, the hammer T1 would not have been able to solve the physical phenomenon of "stress concentration".
Back to smart driving, we also see two routes: Tesla's "all do it" "Apple-like" and other manufacturers' natural choice of "Android-like". Nvidia is a major player on the "Android-like" intelligent driving track.
"Android-like" means that the boundaries between self-development and suppliers will fluctuate. Firmly self-researched like Wei Xiaoli, it will also be considered "OEM".
The hardware and software products released by NVIDIA last night include almost all stages of autonomous driving research and development: sensors, chips, software, maps, algorithms, servers...
Some traditional car companies have been unable to withstand the temptation and have become customers of the "nanny-level program", and car companies that insist on self-research should face similar choices.
What exactly determines the self-development and OEM of intelligent driving? Perhaps there is no answer that everyone is convinced of at this stage.
The only thing that is certain is that the curtain has been faintly opened, and the question is, can you see what is hidden inside?
(End)
Planet article recommendations
"Add Electric Planet Assistant to Join planet community"