Translate the following Chinese Markdown text into English Markdown text, maintaining a professional tone and preserving the HTML tags in Markdown. Only output the result.
Organized by China Electric Vehicle 100 (CEV100) and co-hosted by Tsinghua University, China Society of Automotive Engineers, China Association of Automobile Manufacturers, China Automotive Technology and Research Center, and China Automotive Engineering Research Institute, the China Electric Vehicle 100 Forum (2023) opens in Beijing.
The forum has invited representatives from relevant government departments, automotive, energy, transportation, urban, and communication sectors, as well as industry organizations and leading enterprises, to engage in in-depth discussions on a variety of topics including global automotive industry development trends, high-quality development paths for new energy vehicles, Chinese intelligent and connected vehicle development strategies, power battery and other core industry chain supply chain development trends, changes in new generation automotive consumption, automotive and energy collaborative development strategies, new transportation energy security systems, commercial vehicle transformation directions, automotive aftermarket innovation paths, and digital and intelligent manufacturing models for automobiles.
At the 2023 China Electric Vehicle 100 Forum, Yu Kai discussed the following points in his speech:
Each automaker should carefully consider the decision to create their self-driving microchips. The first reason is the immense financial and cost investment into R&D. The second is how to maintain competitiveness continuously. Therefore, choosing between self-research and leveraging others’ innovations is a strategic decision. If your sales expectation does not reach 1 million vehicles, the investment and efficiency may not add up.
Below is the original text of the speech:
Esteemed Chairman Wan Gang, esteemed Chairman Qing Tai, esteemed Minister Miao, distinguished guests, friends, good afternoon!
Today, I want to share some of Horizon’s insights into the current status and future of autonomous driving computation, especially from the perspective of customer value. I aim to share relatively rational and calm thoughts that do not lose a sense of tension and imagination for the future.
Firstly, I will address a current hot issue, undoubtedly the rapid rise of ChatGPT. On the other hand, the development of autonomous driving does not seem as fast. Why is this?
Secondly, I will share Horizon’s latest developments in technology, products, business, and ecology, specifically recalling what we have accomplished in the last year since participating in the CEV100 Annual Conference.
Thirdly, I will share my personal thoughts on the future.
Without a doubt, ChatGPT can be considered the most prominent progress in artificial intelligence up to this point. I have worked in the AI field for 27 years, and though I have dealt with many language models myself, achieving such development today surpasses the Turing Test, and it is no longer an issue—to doubt this would be unthinkable.On the other hand, autonomous driving has yet to arrive. However, we see that the actual challenges of these two issues are quite different. For instance, ChatGPT, as Li Xiang mentioned earlier, could replace or enhance white-collar jobs, but the tolerance for errors in these jobs is relatively high. For example, if I ask it to write a speech for a 100-person conference, I can roughly outline the main points, and it would write most of it correctly, even if not perfect, as I can modify it based on its output. Autonomous driving is different, especially in the case of self-driving vehicles where the tolerance for errors might be zero since it’s a matter of life and death, and users have high expectations.
Moreover, let’s talk about computation. Computation for OpenAI and ChatGPT happens in the cloud, where there’s ample energy supply, power supply, and an excellent system. However, if it relies on batteries and vehicle-mounted heat dissipation in cars, this challenge is significant, which means that autonomous driving cannot use such large models or computations. Therefore, we can see that Level 4 (L4) and RoboTaxi did not advance commercially, and in fact, companies like Google, Waymo, and Cruise have had layoffs lately.
For instance, we also saw that Argo AI, which was invested in by Ford and Volkswagen, went bankrupt last year. Ford is now focusing on a company for assisted driving, while Volkswagen invested in Horizon Robotics last year and is working with Horizon on mass production, along with software and hardware system innovation for assisted driving.
The industry seems to be returning to its commercial value, its essence, and the user value. What is the user value? Do users indeed demand self-driving vehicles? Our current survey data shows that 87% of users genuinely want a sense of ease during the driving process, eliminating stress and fatigue.
For example, in Beijing, commuting often takes an hour each way. It’s against human nature to focus intently for a whole hour without even glancing at WeChat. People are inherently multitaskers, easily distracted, and not focused. Including an experience I had last year when I traveled from Shandong to Beijing during the pandemic, I queued for about 5 hours because of the epidemic control measures. However, I was driving a Li Xiang ONE vehicle, and during the traffic jam, I didn’t need to keep my foot on the pedal, which made the 5-hour wait much more comfortable compared to situations without assisted driving.In fact, we don’t need to fully achieve autonomous driving; we have already started creating value for users with advanced driver assistance systems (ADAS). As we discussed today, intelligent electric vehicles provide our industry with a chance to overtake competitors. Previously, consumers didn’t recognize Chinese fuel vehicles as high-end brands; however, they do now with intelligent electric vehicles because we have achieved global leadership in intelligent electric technology, especially autonomous driving – a vital factor in purchasing a new car.
So, how is our industry responding to this consumer demand? In Japan and Europe, ADAS primarily focuses on L1 and L2 technologies driven by safety regulations, such as Automatic Emergency Braking (AEB) and Lane Centering Control (LCC). But in China, the situation is different. Safety is the minimum requirement, with a strong focus on user experience and value. Last year, China began mass-producing L2+ High-speed Navigation on Autopilot (NOA) systems, utilizing over 10 cameras, including Surround View systems, millimeter-wave radar, etc.
To my knowledge, the industry’s top-level performance can achieve autopilot takeover at around 100 kilometers. For instance, an investor recently told me about a colleague who drove to a ski resort in Zhangjiakou using our Journey 5 chipset. The vehicle traveled 200 km on the highway without any takeover, making the driver very happy with our investment. However, after leaving the highway and continuing with autonomous mode, they had an accident. This year, leading automakers including Weilai will introduce urban NOA systems (L2++), but I believe there are technological challenges ahead, requiring at least three more years of research to see significant improvements. Currently, the systems require takeover every 20-30 kilometers.
Many automotive executives are present today, and what I’d like to share regarding autonomous driving is not to be overly anxious. Industry development isn’t that fast. From now until 2025, the real focus should be on making NOA systems for highways and closed roads smooth and affordable. At the same time, considerable time and effort are required to bring urban NOA to a usable level. That’s my opinion.Translate the following Markdown Chinese text into English Markdown text in a professional manner, preserving the HTML tags in Markdown and only outputting the result.
By the way, looking at the current consumers, including a media friend I met just now, he mentioned that autonomous driving doesn’t seem as promising as anticipated. For example, we can see that the configuration of autonomous driving, assuming computing power from several tens of teraflops (T) to one thousand teraflops, doesn’t make much difference in user experience and value. What is the reason behind this? Let me try to explain in an engineer’s language. The horizontal axis represents the logarithmic computing power, and the vertical axis represents the value it offers to users. From a few teraflops to 10 teraflops, it’s basically the perception of front cameras, which is the typical L1 and L2, referring to level 1 and 2 assisted driving.
However, from several tens of teraflops to several hundred teraflops or even one thousand teraflops, we find that they all realize high-speed NOA with not much difference, and there is still much work to be done. The dashed line represents the upper limit of value that can be brought to users given our computing power, assuming our algorithms are perfect, our data is sufficient, and our engineering is well executed. What we do is to continuously optimize software and algorithms on a given computing power, and approach the upper limit with more data. We believe that in the coming years, we will reach such a level where a few hundred teraflops can achieve competent NOA in urban areas, but truly realizing autonomous driving in a broader range of areas will indeed require more than a thousand teraflops of computing power.
Regarding Horizon Robotics, we’ve made continuous commercial progress in the past year, including mass production on more than 50 car models, over 120 front-loading models, and close to 3 million automotive-grade autonomous driving chips shipped. Our Journey 5 chip is now one of the two mass-produced chips in the industry with over 100T computing power. In addition, we have secured a series of benchmark vehicle models and automaker mass production projects. Furthermore, working with companies like Ideal Motors, the L8 and L7 models have been delivered since November last year. Last year, we were fortunate to establish a heavyweight strategic cooperation with Volkswagen, a joint venture brand, and we believe we will break through to more international brands in the future.
This figure shows the recent research data from Gao Gong Intelligence, as last year was the first year of mass production for L2+ advanced assisted driving. We achieved the highest market share of 49%, with Horizon Robotics and NVIDIA accounting for 95% of the market share together. As for Horizon Robotics, a startup company established less than 8 years ago, the current progress has been quite satisfactory, and we are grateful for the trust and support from partners and automakers in the industry.Achieving such a commercial breakthrough is actually the result of numerous unseen efforts in technology, safety, innovation, process, systems, and quality. For instance, I’d like to share with you the sense of collecting all seven Dragon Balls in terms of securing certifications in chip safety, architectural safety, toolchain safety, information safety, and network safety. We have obtained top-level safety certifications worldwide, making Horizon’s Journey 5 chip a product designed in accordance with the highest industry safety standards.
Furthermore, without a deep understanding and knowledge of AI software algorithms, it would be impossible to design highly efficient AI chips. As a testament to this, in the 2020 Google Waymo First Annual Autonomous Driving Algorithm Competition, out of 120 participating teams from around the world in 5 competitions, we achieved global championships in 4 categories and second place in the 5th category.
We have not become complacent. Just last week, at the world’s top-level AI computation computer vision CDPR conference, we submitted a paper as first author proposing an end-to-end automatic algorithm framework based on Transformer. This paper was selected as one of the Top 12 best paper candidates among 9,000 submitted papers, showcasing our continuous forward-looking research in software algorithms. This paper is the first to complete a simple architecture for detection, tracking, prediction, mapping, and trajectory prediction using a single neural network end-to-end. The traditional approach divides these tasks into separate modules. This allows us to potentially train the entire autonomous driving system on large-scale data end-to-end, similar to ChatGPT.
Based on this algorithmic understanding, we have integrated forward-looking algorithmic research into chip architecture design and development, creating Horizon’s underlying technology called BPU (Brain Processing Unit). We have trademarked BPU, aiming to build a world-class computing architecture akin to GPU in the future. The BPU is designed for high-level autonomous driving, focusing on optimizing the latest deep neural network algorithm computations.
For example, in our Journey 5 chip, we employ the third-generation BPU architecture called Bayesian architecture. This framework is characterized by its efficient support for Transformer computations, such as with the Swin Transformer, an image recognition algorithm that won the prestigious Marr Prize in computer vision in 2021. Compared to competing chips, we achieve higher computational efficiency with lower power consumption. In another Transformer algorithm, DETR, we also attained the industry’s best FPS computation efficiency.Translate the following Markdown Chinese text into English Markdown text, in a professional manner, retaining the HTML tags within Markdown and solely outputting the result.
What’s the next step? Recently, ChatGPT has provided us with great inspiration. We will continue to use big data, even larger data, even larger models, and learn unsupervised human driving attempts, just like you learn from a large amount of unsupervised, unannotated natural text. This is because each driver’s driving control sequence is just like our natural language text. So, what is a language model?
It is the probability of predicting the next word given a text history. Similarly, we give the current traffic environment, a navigation map, and a history of the driver’s driving behaviors, and predict their next driving action. We can acquire learning from a large amount of unsupervised and unannotated behavior to build a regression-based autonomous driving language model, which is the next step.
At the same time, we can see that past continuous experimental data show that as the model’s parameter scale grows, the overall prediction test loss decreases, meaning that the more parameters the system has, the smarter it is. As seen in ChatGPT, GPT-3 has approximately 175 billion parameters, while GPT-4 has almost one trillion parameters. Our human brain has 100 trillion parameters. You might wonder how many parameters cat and dog brains have – 300 million parameters. From 300 million parameters to 100 trillion parameters in the human brain, the scale of parameters determines the level of intelligence. No mysterious magic is involved; humans are intelligent simply because our brain capacity is indeed large.
What is the next-generation computing architecture? We need to build a unified computing architecture for large-parameter Transformers, especially focusing on the computational efficiency and power consumption within this architecture. We will find that parameter scaling may not be the most energy-consuming part of the actual computation, but rather data storage and data I/O. We need to construct, for instance, a tiered storage architecture to optimize the bandwidth bottleneck under large parameters so that the true computational efficiency can be enhanced while supporting the large-parameter Transformer within the vehicle’s power consumption.
At last year’s Baihui Forum, I proposed for the first time that Horizon should create a business model more open than open source, which is not only providing OEMs with a black-box chip or software, but also helping them achieve their dreams. Many OEMs have dreams of becoming Apple or Tesla. For instance, Li Bin insists on doing everything by himself, and I think many automakers do the same. Therefore, we assist these automakers in creating their own chips and establishing an ARM+Android business model. I proposed this for the first time last year, and since then, we have locked in one OEM, and we are in talks with another.Finally, let me mention that each automaker should carefully consider the development of their own autonomous driving chips. First, there’s a massive amount of funding and cost for research and development. Second, it’s crucial to maintain continuous competitiveness. Strategically, developing in-house or leveraging external technologies is a choice to make. My basic suggestion is that if your sales expectations are not up to 1 million units, the investment efficiency may not yield a positive return.
At the same time, Horizon Robotics is committed to building a flourishing software ecosystem. As we all know, NVIDIA’s greatness and success come from the extensive software ecosystem it has created based on its CUDA architecture. Since last year, Horizon Robotics has been dedicated to building a software ecosystem for intelligent driving and robot computation. The ecosystem construction team has visited all the software companies working on autonomous driving research and development across China, as well as various robotics startups, effectively rallying support.
Now, as you can see, well-known autonomous driving software companies in the industry, such as Lzhou, based on a single Journey 5 chip, have already secured a vehicle project. And others like Pony.ai and WeRide are actually developing their software solutions on Horizon Robotics’ chips.
This is a panoramic view. We’ve built a comprehensive intelligent driving computational ecology in terms of software vendors and hardware domain control, believing that without a software ecosystem built upon China’s own computing architecture, we cannot grasp the initiative for innovation in autonomous and electric vehicles.
Lastly, let me share a slightly contrarian perspective.
Regarding the long-term outlook of ten years, I might be pessimistic on the realization of L3 and L4 autonomous driving. I recall in 2013 when I first led Baidu’s autonomous driving project, I mentioned in an interview that the future relationship between humans and cars might resemble that of humans and horses. If a horse injures a child or another person, is the horse or the rider responsible? However, horses can “drive” autonomously, which is the approach taken by Tesla and other current mass-produced vehicles.
If automakers were to bear this responsibility, innovation would be stifled, and they would hesitate to invest in research and development. I believe the future relationship between humans and cars will still resemble that of humans and horses. While cars can drive autonomously, humans and vehicles must cooperate. Ultimately, the driver should be held responsible in case of incidents.Translate the following Chinese Markdown text into English Markdown text, in a professional manner, retaining the HTML tags in Markdown, and outputting only the results:
However, on dedicated roads, such as those for vehicle-cloud collaboration in autonomous driving, it is possible to achieve unmanned driving under one condition: no human-driven vehicles are allowed on these roads, and all participating vehicles must be self-driving. I believe that under these circumstances, fully autonomous driving can be realized.
Finally, I’d like to share some thoughts I had last week in my friends’ circle. In the first stage, humans dreamt of robots helping us with various tasks, so we trained machines. Unexpectedly, these machines became so smart that now, as Baidu’s Li Zhenyu mentioned earlier, it’s essentially machines training humans. Because humans must find the right way to communicate with machines so that the machines can produce the desired speech drafts. This itself involves humans adapting to machines, just as TikTok’s algorithms continuously recommend content that keeps users engaged, with machines training humans. This is already happening. However, I am concerned that with the current development of artificial intelligence, machine dominance may occur. Humans might feel happy on the surface due to algorithmic satisfaction, but in reality, we give up more of our thinking, and I fear such a future. In the fourth stage, some people (like my Horizon team) awaken and promote equal rights in AI computing between humans and machines. We shouldn’t allow unilateral transparency where everyone’s data is handed over to a machine that operates as a black box, making its workings uncontrollable for us. What should we do? We must promote distributed, localized AI computing that protects users’ privacy and is transparent and decentralized for humans. How? Consider this: in the future, automotive systems will become distributed energy storage centers. I envision that with millions of vehicles fitted with 1,000 T computing power chips, what would they be doing when parked? When I was at Baidu, building a data center with 1 million servers in Inner Mongolia was considered a huge server center. Now imagine billions of cars parked, each equipped with thousands of T of computing power that represents the world’s largest computing resource pool. Therefore, it will certainly provide computing resources for other applications beyond vehicular computing.
With endless possibilities, the computer revolution has just begun. I have also had many discussions with Wang Jun of Changan about the prospects for distributed computing, so I believe such a development is on the horizon. That’s all I have to say. Thank you all!Please provide the Chinese Markdown text you would like translated into English Markdown, and I will be happy to help you. Make sure to include the HTML tags within the text, as requested.
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.