Tesla may well be the biggest driver of China’s journey towards technological self-reliance, as the “Tech Cold War” between the U.S. and China deepens. This observation wasn’t obvious to me until I listened (three times) to Elon Musk’s Q&A session at the 2020 World Artificial Intelligence Conference in Shanghai earlier this month.
While most of the media coverage focused on Musk’s bold proclamation that Tesla is very close to achieving Level-5 autonomous driving capabilities, the entirety of what he said during the Q&A has far-reaching implications, technologically and geopolitically.
Local Hire, Original Engineering, Open Patents
Musk’s Q&A was, first and foremost, an aggressive recruiting pitch for the engineering team that Tesla is building in China. What’s most noteworthy is his emphasis that the Tesla China team will be doing “original engineering”.
This emphasis is important for three reasons:
1. Local optimization: for self-driving to work well, where your engineers are physically located play a surprisingly important role. Musk shared that Tesla’s Autopilot works best in California, because the company’s engineers are mostly in California, thus they can do local optimization most effectively to adapt to the reality in California. The plan is: once self-driving works well in California, the software can be adapted to other locations with varying degrees of further local optimization.
In order for Tesla's Autopilot to work well in China, it needs a local engineering team capable of what its team is doing in California. Thus, “original engineering” makes perfect product sense if Tesla wants to deliver self-driving cars in China -- a growing electric vehicle market where Tesla is taking a large chunk, according to research by ARK Investment. Beyond the engineering and product rationale, it is also a killer recruiting pitch for the best and brightest; no talented engineer wants to spend the day refactoring and adapting someone else’s code.
2. Friendlier regulators: In his Q&A, Musk’s confidence in achieving L5 is tempered by the reality that neither he nor anyone else knows what is the level of safety that must be achieved to convince regulators. Twice as safe as a human driver? 10-times? 100-times? Or in Musk’s words, how many 9’s after 99% is needed? We don’t know.
Regulators play a decisive role in the pace of self-driving advancement, more so than most other tech sectors. For a regulator to make a judgment on the safety threshold, it’s a combination of hard evidence and political will plus incentives. Assuming we have the same safety evidence, then the Chinese government currently has more appetite to experiment with and invest in AI than the U.S. government. Thus, Chinese regulators of all levels are more incentivized to say “yes” to things, while their American counterparts are more incentivized to say “no”. This political will was on full display at this AI conference, where the Party Committee Secretary of Shanghai, Li Qiang, more or less presided over the entire keynote session, where almost every speaker ritualistically praised his leadership at the beginning of each talk. Li’s presence is a big deal, as Shanghai tries to lead China’s AI future, having lost out on the other tech trend booms to Beijing, Shenzhen, and Hangzhou. Tesla’s China GigaFactory is already located in Shanghai. It’s likely that Tesla will continue to receive more favorable regulatory treatment in Shanghai than in California or anywhere else in the U.S.
3. Reduced risk of sanctions: With “original engineering”, Tesla China creates original IP that will not be subjected to future U.S. sanctions. Thus, this approach is effective for recruiting and important for Tesla’s own long-term development. China, in turn, will rightfully have access to the IP generated from Tesla China.
But what’s more important is that Tesla doesn’t care about patents anyways. In 2014, it decided to open source its entire patent portfolio related to building an electric vehicle. In Musk’s mind, patent protection is an impediment not accelerant to technological advancement. In his blog post announcing the no-patent decision, he wrote:
“Technology leadership is not defined by patents, which history has repeatedly shown to be small protection indeed against a determined competitor, but rather by the ability of a company to attract and motivate the world’s most talented engineers. We believe that applying the open source philosophy to our patents will strengthen rather than diminish Tesla’s position in this regard.” [Bold emphasis mine]
Tesla is flexing its muscle to attract and motivate the most talented engineers in China. When they do develop the self-driving technology needed for Tesla (and yes, I believe it is a “when” not “if”), chances are all that technology will be open sourced for anyone in China and the world to do the same. Open source technology, as the Linux Foundation clarified in its whitepaper earlier this month, is not subjected to the U.S. government’s various export control restrictions.
Of course, achieving self-driving is not just an algorithm and software engineering challenge, but also a hardware one. To understand that piece of the puzzle, let’s look at Tesla’s custom-designed Full Self-Driving (FSD) chip in the context of China’s current chip manufacturing capabilities.
What is FSD?
Tesla began developing its own FSD chip in 2016. In 2019, it began volume production and deployed these FSD chips inside Model X, S, and 3. It is an 8-bit chip (not the more powerful 32-bit or 64-bit general purpose chip), in order to satisfy the specific computation needs of a self-driving car but nothing more. Thus far, Samsung manufactures the chips using its 14nm process at its Austin, Texas foundry. A self-driving chip does not have as high a requirement in terms of form factor (or size) as a smartphone or smaller device, thus the 14nm process is sufficient.
The FSD chip is highly optimized to do one particular mathematical operation very efficiently: dot product. It’s how you do matrix multiplications -- an operation that a neural network needs to do many many times, especially for a convolutional neural network. In order for a self-driving car to “see” what’s going on while driving, it has to do a lot of dot products to compute all the image data gathered from its surrounding, frame by frame, before deciding what to do, if anything. And the car must be able to do this constantly for a long time without overheating or slowing down performance. The FSD’s performance apparently exceeds an Nvidia alternative’s by 21-times, while consuming much less power.
Each Tesla car is installed with two of these chips to add redundancy and improve decision-making: if the two chips “disagree”, the decision is thrown out and the computer starts a new compute and inference cycle on new data. Essentially, each Tesla has two “brains” to keep each other honest.
Can China Make FSD?
As I’ve described in a previous deep dive post, “RISC-V, China, Nightingales”, China’s largest chip foundry SMIC does have 14nm capabilities. In fact, SMIC’s co-CEO Liang Mong-song, who is credited with bringing the 14nm capabilities to the foundry, was a TSMC veteran and also instrumental in helping Samsung first achieve its own 14nm process, which is being used to manufacture the FSD chips right now. These chips are designed based on the ARMv8.0-A instruction set architecture (ISA), which isn’t subject to U.S. sanctions since ARM is a UK company.
That being said, the UK recently banned Huawei completely, reversing its earlier decision that allowed Huawei to build the “non-core” portions of its 5G network. This decision may signal a stronger willingness by UK regulators to either push out Chinese companies or prevent technologies from going to China, as it moves closer to the U.S. in this Tech Cold War.
To further muddle the picture, Softbank who owns ARM is mulling a sale after buying it for $32 billion USD four years ago. Meanwhile, RISC-V, the open source ISA, is not yet mature enough to design a self-driving optimized chip like the FSD
Needless to say, there are many moving pieces to keep track of. SMIC is facing many challenges as it grows into a world-class chip foundry. It has had trouble acquiring advanced lithography technology from the Dutch company ASML, due to pressure from the Trump administration. Without ASML’s technology, SMIC will have a hard time advancing beyond the 14nm process to 7nm or denser processes. It will also have trouble acquiring other EDA tools to scale its production capacity, because the dominant suppliers are mostly American. Depending on who you ask, SMIC’s future is still pretty bleak, even with its successful IPO on the Shanghai STAR market.
But at the end of the day, Tesla needs 14nm capabilities and SMIC has them. In fact, it would make sense for Tesla to be a customer of China-made chips for its China-made cars. This arrangement will help Tesla avoid future restrictions if U.S. sanctions widen to include its current chip production in Austin. Producing FSD chips at scale for a high-caliber customer like Tesla will push SMIC to up its game on manufacturing quality and operational efficiencies. Combined with the software advancement in AI from the local engineering team, Tesla can end up giving China’s technology ecosystem quite a boost.
Ambition and Pragmatism
Tesla is not only an ambitious company, it’s also a pragmatic one. Ambitious pragmatism is what a company, and I’d argue a country, needs to both survive and thrive. In the last Cold War between the U.S. and U.S.S.R., there was actually a lot of ambitious pragmatism that catalyzed technological advancements, like rocketry, satellites, and the space program. So far, the current conflict between the U.S. and China looks totally different.
In fact, this “Cold War” comparison between the now and then is a misleading shorthand.
During a time when U.S.-China relations is dominated by ideologues, not pragmatists, Tesla is expanding its Shanghai GigaFactory and aggressively recruiting Chinese AI engineers. Musk is just executing on his own view that “China is the future” and thus an important place to invest in for Tesla to thrive, something he expressed during last year’s World Artificial Intelligence Conference in Shanghai. He’s being ambitiously pragmatic, as an entrepreneur, a technologist, a builder should be.
In the end, California may still be the first place to have L5 self-driving Teslas on the road, thanks to a big head start. But Shanghai won’t be far behind.
2016年，特斯拉开始开发自创的FSD芯片。2019年，开始量产，并在Model X、S和3中都部署了FSD芯片。它是一个8-bit芯片（不是更强大的32-bit或64-bit通用芯片），仅以满足无人驾驶的特定计算需求为目的。目前大批量生产FSD芯片的厂商是三星在德克萨斯州奥斯汀的铸造厂，使用的是14nm制造工艺。 因为驱动无人驾驶的芯片在外形或尺寸方面没有智能手机或更小的设备要求那么高，所以14nm制造工艺就足够了。
FSD芯片的特点是高度优化一种非常具体的数学运算：点积。这就是做矩阵乘法的方法，也是一个神经网络（neural network）需要做很多次的一种运算，尤其是“卷积神经网络”（convolutional neural network）。一款无人车在开的时候，要想决定自己下一步该做什么，就得先把它周围所吸收的所有图像，一框接一框的做出计算，而且必须能够长时间持续做这些计算，还不会过热或降低性能。FSD的性能超过了英伟达的一款竞品21倍，同时耗电也更少。