An dem Abend, als dieser Chip erstmals eingeschaltet wurde, brachen die chinesischen Ingenieure alle in Tränen aus.

Early spring 2024, Beijing.

The air was filled with the smell of adrenaline - not at a drag - racing site, but a mix of 53 - degree Moutai and engineers' tears.

When the TX81 chip was first powered on, Wang Bo and his team were too excited to contain themselves because at that moment, they might have rewritten the history of Chinese technology.

01 Chinese Rewrite Moore's Law

A year later, DeepSeek emerged out of nowhere, cutting through the Western technological arrogance like a sharp blade.

Just as cloud giants quickly followed suit and connected to DeepSeek one after another, a startup, Qingwei Intelligence, quietly joined this technological wave.

Although it had been established for less than seven years, its solution was highly disruptive.

Because in Qingwei's servers, there was an unusual heart beating - the TX81 chip.

Over the past few decades, the bet on computing power improvement was placed on the process race.

However, the physical limit of Moore's Law was approaching rapidly. The growth curve of traditional chips seemed pale and sluggish in the face of the explosive computing power demand of AI.

Moreover, overseas giants had built strict barriers in technology and ecosystem. For China's computing power to break through, there was only one way - architectural innovation is the ultimate password to break the wall.

The TX81 chip hit the rhythm of this historical moment.

This is a dynamic reconfigurable chip (RPU) with a brand - new architecture.

It can perform efficient calculations without the complex exchange between the instruction set and storage. According to different scenarios and tasks, such as voice, image, and large models, it can dynamically match computing units, forging the precious computing power generated by mature processes into a "Transformer" that can meet different computing needs.

Thus, it has forged four core advantages of high energy efficiency, high concurrency, high scalability, and high cost - effectiveness.

It can be said that this is a chip "native" to the AI era.

China's computing power chips started later than those in the West. The call for "domestic substitution" has been sounded for many years. However, different from traditional domestic substitution, Qingwei is taking a perilous new path of high - level domestic substitution:

Starting from scratch is extremely difficult, but it is related to the autonomy and lifeline of China's AI computing power foundation.

This is similar to the counter - attack script of China's automobile industry.

In the past few decades, core technologies such as engines and gearboxes were monopolized by foreign enterprises. Domestic fuel - powered cars were "crushed" by them.

As latecomers, how could they break through and even overtake? It was almost impossible to follow in the footsteps of the existing model and make minor repairs.

Domestic automobile brands such as BYD, Wenjie, and Xiaomi achieved "overtaking on a new track" through electrification, intelligence, and networking. In particular, key technologies such as batteries, intelligent driving, and integrated casting have even reached the world's leading level.

Relying on a huge market, industrial chain integration, and favorable policies, domestic electric cars offer high cost - effectiveness, allowing consumers to get the configuration and experience of a million - dollar foreign luxury car for just a few hundred thousand yuan.

The leap of China's automobile industry has not only reshaped the global automobile landscape but also become an excellent model for industrial upgrading of late - developing countries.

Qingwei's journey in the computing power industry is exactly the same as the overtaking on a new track of domestic electric cars.

It is using a series of original and leading "combination punches" such as reconfigurable computing, C2C computing power grid, and 3D storage stacking to open a new era of AI computing power.

This is a Chinese path of innovation from the bottom up, and the entire chain is firmly in the hands of the Chinese people themselves.

02 Entrepreneurship Means Going All In

Qingwei's birth began with a sudden flash of inspiration during a "technological arduous journey".

In 2018, Wang Bo, who was working on machine vision in a technology company, hit a wall when selecting chips.

At that time, he was choosing a chip for his product. After searching around, he found that those well - known chips in the industry did not meet expectations: they were expensive, had high power consumption, and poor heat dissipation.

In his anxiety, Wang Bo saw Professor Yin Shouyi from Tsinghua University's THINKER reconfigurable chip on his WeChat Moments.

This piece of news was like a bolt of lightning splitting through the fog. His intuition told him that this was a "new species" completely different from existing products.

The next day, on the Tsinghua University campus, Wang Bo and Professor Yin hit it off immediately: Wang Bo would invest money, and Professor Yin would be responsible for bringing the laboratory results to the industry.

However, there was a huge gap between the ivory tower of scientific research and the battlefield of the industry. Universities are good at frontier exploration, but there are still a lot of complex engineering and supply - chain links in product development.

As a result, it took more than half a year, and finally ended in failure.

But failure ignited the stubbornness in Wang Bo's bones.

After much thought, he resolutely decided to burn his boats and founded Qingwei Intelligence with Professor Yin, vowing to turn the reconfigurable chip from a paper concept into reality. For this, he bid farewell to the glory of a listed company.

What supported his all - out effort was his firm belief in the future of reconfigurable chips.

In the AI era, computing power is as necessary as water and electricity, and the demand is everywhere.

The path of traditional chips to improve computing power through the process race is gradually reaching a "dead end". However, the reconfigurable architecture is like a glimmer of light breaking through the iron curtain.

"If this succeeds, it will be enough to leave a name in the history of Chinese technology."

This thought made Wang Bo excited, but he knew well that the road to the summit was as perilous as climbing Mount Everest.

Although the Americans proposed the concept of reconfigurability in the 1960s and the Tsinghua team had been working on it for more than a decade.

No one in the world had ever tried to turn academic sparks into a commercial wildfire.

This was exactly the biggest challenge.

At that time, AI computing power chips with GPU architecture monopolized more than 90% of the domestic market. Wang Bo's team was in a weak position.

At the beginning, more than twenty people were crowded into an old office building in Zhongguancun, and the office area was only 30 square meters.

Among these people, there were many technical experts who had jumped from large chip companies.

Previously, they worked in high - end 4A office buildings, but now they were in a "pigeon - cage" of a startup. The working environment was poles apart.

For the first few months, the company didn't even hang a sign. Passers - by once suspected it was a fraud company and even reported it to the property management.

In such a "humble abode", without desks and chairs, they bought them on Taobao and assembled them by themselves. Two long white tables and more than twenty simple office chairs were Qingwei's first "trenches".

Everything started from scratch. The precise and profound physical design in the computer, NPU architecture, internal module research, PV, packaging... Even the chip R & D process and quality standards needed to be built one by one.

Yang Feng (a pseudonym), one of Qingwei's first - batch engineers, recalled that when the technical director drew the architecture diagram of the reconfigurable chip on the small blackboard, he was extremely excited:

"When we thought about doing something unique in the industry, there was a fire burning in everyone's pupils."

Despite the difficult conditions and the uncertain future, the team's entrepreneurial passion was just like that of those great companies when they started in a garage.

This team is bravely challenging a pioneering cause.

Under the wave of domestic substitution, a sense of mission and responsibility is an ever - shining lighthouse in the dark night. Even though the conditions were so poor that they were ashamed to show others and the project needed to be kept secret so they couldn't share it with those around them.

But faith, like a lighthouse in the dark night, supported them on this long and difficult journey.

03 The Difficulty Is Comparable to Climbing the North Slope of Mount Everest

Creating a new - architecture chip from scratch is already a perilous peak, but Wang Bo said that Qingwei can only truly enter the first - tier of the industry by making AI computing power chips.

To achieve this goal, the team decided to fight to the last.

AI computing power requires a larger chip area to accommodate more computing units. Qingwei chose a single die size of 800 square millimeters, which is almost the physical limit of chip size.

In addition to the challenge of chip area, they also need to use reconfigurable technology to improve the computing power utilization rate to a new height. This not only involves a single card but also requires good communication between cards.

Since there was no precedent to follow, they had to feel their way forward.

Regarding multi - card interconnection and better programming, after research, the team proposed a bold technical path - C2C grid direct connection.

Different from the traditional model, C2C grid direct connection does not require a switch, which means cutting off the data transfer station. It not only saves costs but also greatly reduces the amount of data congestion, allowing the chip to unleash surging computing power.

To overcome this hurdle, they need to build a scheme model, conduct simulation verification, and find components with extremely stable performance.

Although the team had some experience before, in the field of large - computing - power chips, there was not a single line of ready - made code. Every line of code was original. Moreover, at that time, the shadow of the pandemic loomed.

During this period, the team encountered an unexpected technical problem: to give full play to the advantages of the reconfigurable chip, they needed to customize a clock device.

The clock device is the basis for the chip's clock operation and is related to the chip's performance.

Tom (a pseudonym), a chip back - end technical expert who had worked at Huawei and Alibaba for many years, volunteered to take on this task.

Although physical implementation was Tom's specialty, he didn't have mature experience in this field. "At this critical moment, I had to bite the bullet and go ahead."

Since this link was very critical and even related to the success of the final tape - out. "The cost of a single tape - out is as high as tens of millions of US dollars!" Tom joked: "If it fails, we can all look for jobs together with me."

At that time, Tom and his colleagues had fevers one after another. According to his recollection, one day when he woke up, he suddenly felt extremely cold, with aches in his bones, and his body temperature soared to 39.2℃.

Ibuprofen, a fever - reducing medicine, was out of stock. Taking advantage of the last bit of strength, he rode a shared bicycle back to the company and grabbed several packs of Ganmaoling from the office medicine box. When he got home, he locked himself in the study and continued to tackle the problem despite the pain.

When encountering difficult problems, Tom would search through global literature, learn, borrow, and adjust... Repeating this process. Under high pressure, he, who never had acne before, suddenly had pimples all over his face during that period.

Choosing a new path is destined to be full of thorns.

The problem of NPU operator switching control logic held up the team for half a month; the chip routing problem made the whole team work hard for more than a month... The iteration of a mature path usually takes about 12 - 18 months, while the R & D cycle of an original new architecture may be twice as long.

When sending the chip for tape - out, Yang Feng (a pseudonym), an engineer, said, "My heart was in my mouth. If there was a single mistake in the process, all our efforts over such a long time would be in vain."

On the night when the tape - out was sent back, the atmosphere was tense. At eleven o'clock at night, Yang Feng brought the chip into the office and saw CEO Wang Bo, CTO Ouyang Peng, and all the technical backbones holding their breath and waiting.

However, after two hours of power - on debugging, the chip still didn't light up. The meeting room began to stir, and everyone's faces became gloomy.

After a thorough check, it was found that the problem was the matching between the circuit board power supply and the chip.

After debugging, the chip finally lit up. Looking at the clock, it was already two o'clock in the morning. Wang Bo and Ouyang Peng took out the prepared Moutai and filled everyone's glasses.

That night in Beijing, the air was filled with the smell of adrenaline - not at a drag - racing site, but a mix of 53 - degree Moutai and everyone's tears.

04 This Time, He Will Never Retreat!

Countless giants have emerged in the global history of technology.

These giants have led one technological era after another. However, no matter how powerful these giants are, there are always latecomers who can overthrow the predecessors.

The reason why these latecomers can achieve a successful counter - attack is not by copying but by doing different things in the historical trend.

Twenty years ago, when internal combustion engines still dominated the world, Elon Musk, following the trend of new energy, used Tesla's electric motors to subvert the century - old automobile industry.

Doing different things is the most prominent label on Wang Bo and his team.

Wang Bo, with a technical background, participated in the research and development of China's first - generation switches and routers when he was studying communication engineering at Beijing University of Posts and Telecommunications.

At that time, the domestic communication market was monopolized by international giants such as Cisco, and Huawei had just started to develop domestic switches. Wang Bo said that at that time, he and Huawei were comrades - in - arms in the trench of domestic substitution, but unfortunately, he didn't fight side by side with Huawei to the end.

Later, Wang Bo caught up with China's first wave of cloud computing.

In the competition surrounded by large companies and strong competitors, he decisively avoided the public cloud track and bet on the private cloud, winning a place for the company and finally successfully listing on the A - share market.

These two journeys have tempered Wang Bo's profound insights and persistence in the technology industry:

Only by daring to be the first to enter the uncharted territory can one seize the opportunity or achieve a late - comer counter - attack.

The reconfigurable chip is exactly such an "uncharted territory" full of temptation and danger.

This "original dream" not only ignited Wang Bo but also attracted many ambitious young people to join.

"This is not only a highly challenging exploration but also a very cool thing."

After six years of hard work, they made it.

In 2025, Qingwei Intelligence's reconfigurable cloud - based computing power chip TX81 was the first in the world to achieve large - scale mass production of similar chips and several thousand - card cluster computing power centers were established across the country.

Interestingly, Z