Huawei's Ascend chose the hardest road: not compatible with Nvidia
For three or four decades Chinese IT has been a follower. Ascend is Huawei's first product that does not solve a chokehold problem but instead goes into uncharted territory, starting at the same time as Nvidia. Xu Zhijun said: anything you do as a follower gets more exhausting the longer you do it.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Ascend's harshest decision was not being compatible with Nvidia
At Ascend's founding there were two choices. The first was whether AI chips should be specialized or general-purpose; Liao Heng, as chief scientist, judged that AI should be a general-purpose product, and the architecture choice has been proven right today. The second was harder: whether to be compatible with Nvidia. Many domestic chips chose compatibility, with the benefit that Nvidia's existing ecosystem and tools could all be shared. But Xu Zhijun was firm from the start that they must take a different path. Fang Xingdong believes this choice was extremely bold at the time, and that many times along the way they faced enormous challenges — because the entire Chinese IT industry has been a follower industry for three or four decades, and had never developed based on its own root technologies.
— Fang XingdongAnything you do as a follower gets more exhausting the longer you do it
Fang Xingdong recalls having dinner with Xu Zhijun three years ago, when Xu Zhijun said, half to himself: Huawei has done this for so many years, and the discovery is that anything you do as a follower always gets more exhausting the longer you do it. There are two kinds of following: in standardized products, Chinese companies can always catch up and even overtake, but by the time they overtake, there is no profit left in the industry — 2G to 5G, PCs, phones all went this way; in non-standardized areas, such as competing with Cisco in routers for so many years, they have not drawn closer. In Xu Zhijun's eyes, what Huawei did in the past was also following. So Ascend is special among the six products: the other five solve chokehold problems and address survival first, while Ascend was not choked in 2019 — it is a fully proactive compute product.
— Fang XingdongWith process choked off, only system innovation can make it back
Each generation of industry chips gains roughly 20% or more in performance, mainly from the increase in transistor count driven by advanced process improvements. Huawei lost this premise and can only gain performance through system innovation and breakthroughs — stacked compute, interconnect, higher integration. Fang Xingdong explains with a car metaphor: before, you just had to make the engine stronger and stronger; now the engine is constrained and cannot be enlarged, so you can only make the car run well through architecture, weight reduction, heat dissipation and other factors. Nvidia's solution is to keep working on the engine, because TSMC can still manufacture it. In the interview Liao Heng uses the metaphor of an 18-story pagoda: many companies are good at only one layer because of the great division of labor, while Huawei is forced to do manufacturing, ecosystem and chip design all at once.
— Fang XingdongHuawei has to pull the domestic supply chain up with it
Previously Huawei did design, relied on the global supply chain, had TSMC manufacture, could choose the best international partners for other components, directly buy American products and assemble them, sell at a good price with a brand premium, and profits were decent. Now it cannot cooperate abroad, TSMC cannot process for it, and domestic processes are still lacking in every respect, so Huawei has no choice but to support and drive the domestic supply chain — only then does it have a way to survive and get products manufactured. So the challenge of the past seven years is not just improving its own capabilities, but also driving the entire supply chain — without driving it, independent autonomy is impossible. Manufacturing is not Huawei's main business, but it has to help companies like SMIC and Yangtze Memory keep advancing process, yield and quality.
— Fang XingdongAscend's significance is not that Huawei survived, but that the industry has an engine
In the past, Huawei's product launches and iterations were still based on American root technologies and the industrial ecosystem, and the high-value components of a phone were basically foreign. Now Huawei's phones, including Ascend, almost 100% drive the domestic supply chain. Fang Xingdong believes the greatest significance of the Tao Law is not that Huawei itself can survive and be autonomous, but that China's entire semiconductor industry and entire IT industry have their own engine, which never existed before. Huawei's rapid development represents the rapid development of the entire industrial ecosystem. He also points out that the market, looking at reports, news or public opinion, has not yet realized that we are at this inflection point in the cycle — mention Nvidia and people think it is formidable, mention Huawei and they say it is policy that makes everyone use it.
— Fang XingdongIt is not a question of difficulty, but of cognition and resolve
Fang Xingdong believes doing communications versus doing IT is essentially not a question of difficulty but of cognition and resolve — whether you dare to do it. The most typical example is the Gauss database: among these six products Ascend stood out, and the least conspicuous may be Gauss, but the database is a half-century-old industry with a relatively stable market, and if it really gets choked it is deadly. Huawei at the time wanted to acquire or support domestic vendors, but after research found none of them were up to it; domestic companies could only grow in the cracks of American giants, with small scale, weak R&D capability and weak product innovation. So Huawei decided to do it itself, benchmarking Oracle, and from 2011 and 2012 invested 1 billion, 2 billion, 3 billion yuan a year, without knowing whether there would be any output in the future.
— Fang XingdongThe future is two centers, not one replacing the other
Fang Xingdong judges that global high tech moving from a single system centered on the US to two centers is the general trend; politics will not allow a continued single center — the US will not allow it, and China will not allow it either. He believes a single center is itself abnormal, and only two centers can create competition — only with two systems does every consumer, every company, every country, no matter how big or small, have a choice, and not be choked into a forced either-or. So the mission of Huawei's six products is to go from being the second choice to being the better choice; it is not a simple substitution relationship with Nvidia, but each has its strengths, they complement each other, and they compete and cooperate. He admits the American system is currently stronger, with its advantage mainly in advanced manufacturing, and that Huawei, under process constraints, relies on the Tao Law not to be left behind; once advanced process is also broken through, it can work both ways at once.
— Fang XingdongThe tech war won in the short term, at a greater long-term cost
Fang Xingdong believes the US did indeed suppress Huawei through the tech war, but cannot solve the problem in the long term. He reasons: without the national confrontation of 2017 and 2018, Huawei might have been at least one-third of Nvidia and one-third of Apple, possibly China's largest company by market cap; in 2019 Huawei had already surpassed Apple, and its global shipments briefly surpassed Samsung for one quarter. Nvidia today is dominant alone; if it had from the start advanced side by side with Huawei in AI chips, it could not have reached today's position. He judges that once the side being attacked builds a self-cycle that does not rely on the outside, and the other side cannot eliminate it even under the harshest conditions, it will be even harder to eliminate later; it will only grow stronger, turning from one monopoly into two competitors.
— Fang XingdongIn their own words · checked verbatim
He said Huawei has done this for so many years, and the discovery is that anything you do as a follower always gets more exhausting the longer you do it.
他就说华为做了这么多年，这个发现就是，但凡做追随的事情，总是这个越做越累。
Fang Xingdong19:11
I have been in this industry for 30 years, and I feel that many of my past understandings can all be reset to zero.
我在这个产业里面30年，我觉得原来过去的很多认知，我觉得都可以给它清零了。
Fang Xingdong22:11
Without the tech war, Huawei is an excellent company, but it might not, would not become a great company; but with the tech war, it will certainly become a great company.
没有科技战，华为是个优秀的公司，但是他可能不是，不会成为一个伟大的公司，但有了科技战以后，他就一定会成为一个伟大的公司。
Fang Xingdong52:52
So the hurdles Huawei currently faces are, with technology, actually about driving the improvement of all of China's industry, the entire supply chain, but many things it has to do itself first, because if it does not do them, no one else can.
所以华为目前这几关，跟技术，实际上就是它带动的是，整个中国的这个产业，整个这个产业链的这个提升，但是很多事情，它不得不自己先干，因为它不干，它就没能干。
Fang Xingdong1:00:58
Once the side being attacked, whether Huawei or China, builds a self-cycle, a self-cycle that does not rely on the outside, frankly speaking it becomes a long-term competitor.
一旦被打击的一方，比如说华为也好，或中国也好，他构建了自循环，就是不靠外界的自循环，坦白讲他就变成了长期的竞争者。
Fang Xingdong1:22:13
Figures
| Year Ascend was founded | 2017 | 15:08 |
| Original planned completion time for Huawei's six-book project | Five years, to 2024 | 2:02 |
| Originally planned number of interviewees for the six books | 50 per book, about 300 in total | 3:02 |
| Actual total interviews | Estimated to exceed 500 by the end of next year | 4:03 |
| Performance improvement per generation of industry chips | About 20% or more | 29:16 |
| Ascend 950 mass production | The second half of this year | 26:12 |
| Huawei's 2019 shipments | Briefly surpassed Samsung for one quarter to become global number one | 1:20:12 |
| Nvidia's current market cap | $5 trillion | 1:20:12 |
Glossary
- CANN / Huawei AI computing framework
- Huawei's AI software ecosystem and operator development framework, benchmarked against Nvidia's CUDA.
- Tao Law / system-level performance improvement path
- When process is constrained, trading architecture, interconnect, packaging and other system optimizations for chip performance gains.
- Red Army Blue Army / Huawei's internal adversarial decision-making mechanism
- An internal mechanism that sets up pro and con sides to debate the same decision, used for strategy and solution validation.
- Software-hardware co-design / joint chip-model design
- Compute companies and model companies communicate in advance about next-generation features and jointly define products.
How to listen
Founders, investors and engineers watching domestic compute, the semiconductor supply chain and AI infrastructure, especially those who want to understand the difference between Ascend's and Nvidia's routes.
The first ~13 minutes cover the origins of six books and the interview process, with little bearing on investment judgment.