AI芯片全新架构，如如AI打造极致性价比
In the era of large models, AI entrepreneurship is flowing into various industries. Among them, the computing power field is regarded as one of the best opportunities for AI entrepreneurship.
With the rapid development of large models, computing power has become the core driving force in the field of artificial intelligence. From GPT-4 to the latest super-large-scale models, the training and reasoning requirements behind them are increasing at an unprecedented speed. The demand for computing power directly promotes the research and development of AI-specific chips, thereby driving the enthusiasm for chip entrepreneurship.
In China, capital continues to enter the chip market, and many AI chip startups are constantly emerging. RuRu AI is a startup focused on creating large-model reasoning chips. With a team with strong technical strength, it has attracted a lot of attention from the outside world.
In the recently held second season of "Lenovo's New Business Innovation Ecological Roadshow", RuRu AI was selected as one of the TOP 10 projects, becoming a new star in Lenovo's ecosystem. The selection of RuRu AI not only demonstrates its own technical strength but also responds to Lenovo's SMB Business Group's long-term commitment to startups: to accompany innovative enterprises to grow with all-round, high-quality, and one-stop services. Lenovo's SMB Business Group is willing to work hand in hand with customers and witness their success.
"The era of large models will change everything," said Huang Xiaoyu, the founder of RuRu AI. "Customized dedicated chips for generative AI are about to become a trillion-dollar blue ocean market. High-performance and low-cost reasoning computing power will be the core demand of users."
A New Architecture with Extreme Cost Performance
Under the background of the intensified global AI competition, China is accelerating the construction of computing power infrastructure to consolidate its competitiveness in the era of large models. Industry insiders believe that the active support of the government and the inclination of relevant policies will further reduce the cost of computing power and improve the R & D efficiency of AI enterprises. However, the entrepreneurship of domestic chips is not achieved overnight, and the difficulties to be faced include:
First, the technical threshold is high, and the R & D cycle is long. Reasoning chips need to find a balance between performance, power consumption, and cost, and innovative chip architectures often require years of R & D investment. Compared with foreign countries, domestic startups have a gap in technical accumulation. Moreover, the construction of the domestic software and hardware ecosystem is insufficient, making it difficult to attract widespread adoption by developers.
Second, the market competition is fierce, and they also face financial and resource pressures. Several giants dominate the AI chip market, and it is difficult for domestic startups to compete with them directly. Chip R & D requires continuous capital investment. From design to tapeout to mass production, a single link may consume tens of millions or even hundreds of millions of yuan. However, the financing ability of startups is limited, making it difficult to support long-term R & D.
Third, there are unstable geopolitical factors. Some core resources for chip development are subject to foreign countries, which directly affects the chip R & D and production process.
In the face of many challenges, RuRu AI can still maintain its own R & D rhythm and move forward steadily, which comes from the solid technical foundation of its core team.
The founder Huang Xiaoyu's strong interest in chips can be traced back to his childhood. It was in the early 1990s when he first came into contact with a computer and was full of curiosity about how the computer works. This curiosity has been driving his academic and professional career. In 2007, Huang Xiaoyu went to the UK to study and became interested in chip design. He specialized in microelectronics and chip design during his postgraduate studies.
After graduating, Huang Xiaoyu stayed in the UK and worked at AMD and Broadcom for more than ten years. His doctoral supervisor is Steve Furber, a fellow of the three academies of the British Academy of Europe, who led the team to create the ARM architecture in Cambridge, UK in 1983, and is therefore known as the "Father of the ARM Architecture".
When ChatGPT emerged in 2022, Steve found Huang Xiaoyu and said that the situation is very similar to when they established the ARM architecture. The new demand for computing power from large models is calling for a new and efficient chip architecture. Therefore, Huang Xiaoyu decided to establish RuRu AI and invited Steve to join as a co-founder.
RuRu AI is committed to creating a Transformer-specific ASIC reasoning chip. There are several technical paths for AI reasoning chips, such as GPU, FPGA, ASIC, etc. Huang Xiaoyu believes that ASIC will become the mainstream development direction of AI chips in the future. The reason is that when the model enters the large-scale application stage, ASIC can perfectly couple with neural network-related algorithms; compared with general-purpose chips, ASIC has less redundancy, low power consumption, high computing performance and efficiency, and the larger the chip shipment, the lower the cost.
Thus, RuRu AI has launched a new Vajra architecture to deal with the high-concurrency scenarios of large-model applications. Huang Xiaoyu introduced that the Vajra architecture can tailor the ASIC architecture for the algorithm model to achieve the ultimate cost performance and energy consumption ratio. At the same time, the Vajra architecture's pioneering large-model architecture reconstruction algorithm realizes reconstruction at the hardware level and can support almost all variants of Transformer models.
"We expect to be able to do some work that GPUs cannot complete. Doing an ASIC such a customized architecture around Transformer can make the customer's reasoning performance infinitely close to the theoretical maximum, thereby reducing costs and increasing efficiency," Huang Xiaoyu said.
At the Roadshow Site, Joy and Assurance
The heat of the chip track is bound to lead to increasingly fierce competition. In this red ocean market where the competition is fierce, the game between big fish and small fish will continue. Huang Xiaoyu believes that compared with those large chip manufacturers, the advantage of startups is that they are not likely to fall into the "Innovator's Dilemma", that is, they may hesitate when entering the emerging market.
"As the saying goes, those who have nothing to lose are not afraid. We can go all in. If we succeed, there will be a huge incremental market. If we fail, we can quickly turn to another direction," Huang Xiaoyu said.
However, although startups are "easy to turn around", when it comes to competing with hard power, it will be found that going it alone may not be the best solution for SMEs. The same is true in chip entrepreneurship. If you can get the support of a stable and powerful ecosystem, it will naturally be more effective with less effort.
It was also in the third year of entrepreneurship that Huang Xiaoyu noticed Lenovo's ecosystem and Lenovo's "Star Plan". He found that Lenovo's investment in the AI large model is very firm and has launched many innovative measures. "There is no doubt that Lenovo's philosophy is very consistent with ours," Huang Xiaoyu said. He signed up for the second season of "Lenovo's New Business Innovation Ecological Roadshow" at the first time.
Before this, due to years of working and living in the UK, Huang Xiaoyu had a very good impression of the Lenovo brand, "because Lenovo is very great in Europe". He observed that European engineers around him have gradually replaced their office machines with Lenovo products in the past ten years. The reason is that the cost performance of Lenovo products has far exceeded some Western top brands.
With the mentality of learning from peers, Huang Xiaoyu carefully made a PPT for this roadshow. At the roadshow site, he met many big names in the industry, "names that were previously seen in public accounts or reports", and he was very excited. During the speech, he also shared the mental journey of doing R & D and the real feedback from customers.
In Huang Xiaoyu's view, Lenovo, as a global leading technology enterprise, has complete industrial chain resources, R & D capabilities, and market coverage, which can provide important support for AI chip R & D. Lenovo has a deep accumulation in the hardware fields such as PCs, servers, and data centers. Its global supply chain resources can support the design, production, and delivery of AI chips, helping startups or partners enter the market more quickly.
And Lenovo's huge ecological network is backed by a strong resource integration capability, which can help the AI chip R & D team obtain more software and hardware adaptation support and reduce the problem of technical isolation.
Therefore, when RuRu AI was selected as one of the TOP 10 projects and entered Lenovo's "Star Plan", in addition to joy, Huang Xiaoyu felt more assured. He felt that the road ahead is no longer lonely.
The Support of the Ecosystem, the Confidence of Entrepreneurship
In his spare time in the industry, Huang Xiaoyu likes to study Buddhist scriptures. "The wisdom of Buddhist scriptures is actually about how the human mind works," Huang Xiaoyu said. "To some extent, what Buddhist scriptures and artificial intelligence want to explore coincides."
Buddhist scriptures have also influenced the company culture of RuRu AI. Both the company name and the product name are inspired by Buddhist scriptures. Huang Xiaoyu said that the name Vajra comes from the "nature of the mind" in Buddhist scriptures, which entrusts the R & D team's expectations for the future: "Perhaps a digital mind will really appear in the future."
The wisdom of Buddhist scriptures is also of great significance to Huang Xiaoyu, "like a bright lamp", giving him full confidence when doing R & D or entrepreneurship.
Now, this confidence has an additional layer, that is, the strong support from Lenovo and Lenovo's ecosystem.
Huang Xiaoyu still remembers that many enterprises came to communicate with him at the roadshow site. A large state-owned bank showed great sincerity. Not only did it invite him to visit the bank's headquarters, but it also brought in the bank's internal technical experts for communication. Venture capital circles and industry-leading institutions have also noticed RuRu AI.
Lenovo is the biggest reason for the emergence of these potential cooperations. Huang Xiaoyu believes that Lenovo's layout in the AI large model is very advanced. They carry out business around AIPC and play an important role in promoting the landing and industry development of domestic large models.
Obviously, Huang Xiaoyu is full of expectations for Lenovo's ecosystem: "We have just started to enter Lenovo's ecosystem. In the future, perhaps our chips can cooperate with Lenovo's business, or with the help of Lenovo's ecological supply chain, we can accelerate our own development."
For RuRu AI, 2025 is a crucial year. After two years of R & D, it is expected that at the end of this year, RuRu AI's self-developed mass-produced chips will be officially taped out. With the help of Lenovo's ecosystem, Huang Xiaoyu believes that the company's development will reach a new level.