Tsinghua PhDs are "selling" computing power and "melting"
Latest update time：2024-09-15
Reads：
Professors lead PhD students to compete in the primary market.
Text丨Lieyun Jingxuan ID: lieyunjingxuan
Author: Sun Yuan
In the AI field, a rising star from the Tsinghua School was born.
Behind a nearly 500 million yuan Series A financing, there are 15 new shareholders. Led by the Social Security Fund Zhongguancun Independent Innovation Special Fund (managed by Legend Capital), Qiming Venture Partners and Hongtai Fund, Lenovo Capital, Xiaomi, and Softbank High-Tech rushed into the market.
Following this round of financing, we will find that 10 months ago, this startup, which was only established for half a year, received "group buying" angel rounds from over 13 institutions, and Baidu, Tencent Zhipu AI were both on the list of shareholders.
Since its establishment, this Tsinghua-affiliated company has been targeted by prominent VCs, including Sequoia Capital, Northern Lights Capital, and Jinshajiang Capital. In addition, Cao Xi, a former partner of Sequoia China, also came to place a bet with Monolith Capital.
According to incomplete statistics from Lieyun.com, this AI startup, which is only 16 months old, has received support from over 30 capitals and has raised nearly 1 billion yuan in total, making it a well-deserved darling of venture capital.
According to relevant investors, the valuation has basically completed three consecutive jumps after three rounds of financing, approaching 10 billion. In other words, after this round of financing, a new AI unicorn will be born.
And it is InfinigenceAI, a computing power operator led by Xia Lixue, a post-90s PhD from Tsinghua University.
In fact, if you look through the information, you will find that Wuwen Xinqiong is always directly equated with AI infrastructure, but Wuwen Xinqiong CEO Xia Lixue emphasizes the role of "operator".
He believes that as the AI field gradually develops, the demand for computing power will form a richer market environment, and the emergence of operators will be necessary.
In his outline, operators, unlike the single-point optimization tools or software-defined AI infrastructure commonly understood in China, can do All in One and bring computing power products as standardized as water, electricity and coal to the market, and supply them to the smart ecosystem for everyone who provides services or applications.
This original intention and vision is also the differentiated approach of Wuwen Xinqiong.
Tsinghua professor born in the 1980s is in charge, and more than 30 VCs come to invest
Behind Wu Wen Xin Qiong, there is a team of nearly 100 Tsinghua PhDs, among whom Xia Lixue’s teacher Wang Yu is the "money-making stone".
According to Tianyancha, Wang Yu is the founder of Wuwen Xinqiong. Born in the 1980s, he has many labels.
Wang Yu entered Tsinghua University in 1998 and completed his undergraduate, master's and doctoral studies. He has been teaching at the university since 2007 and became the dean of the Department of Electronic Engineering at Tsinghua University at the age of 38. Since then, he has been the first post-80s dean of the Department of Electronic Engineering at Tsinghua University and the second youngest dean since the establishment of the department.
Focusing on research in hardware, chips and basic software, Wang Yu has won the best paper award at top conferences many times. He has published more than 50 articles in IEEE/ACM magazines and has been cited more than 17,000 times by Google Scholar.
With good academic performance, one can start his own business. In 2016, Wang Yu entered the hearts of VCs in the primary market.
At that time, the deep learning processor project headed by Wang Yu passed the demonstration and received support from the School of Electronic Information of Tsinghua University. Afterwards, the project team used the research results "Method and Device for Layer-by-layer Variable Precision Fixed Point of Convolutional Neural Network" as a price to invest in AI chip company DeePhi Technology for industrial operation.
In the same month of its registration, DeePhi Technology won a US$5 million angel round from Gaorong Ventures and GSR Ventures. In the second year, its valuation exceeded 1 billion. Subsequently, Ant Group, China Merchants Venture Capital, Samsung Ventures and others "snapped up" the A+ round of financing. Finally, it was acquired by FPGA manufacturer Xilinx for US$300 million two years after its establishment, marking a successful conclusion.
As a result, SenseTime became the first AI chip startup in China to go public, and also the first company to complete the transformation and return loop since Tsinghua University established the technology achievement transformation system.
According to ChinaVenture, first-tier investors have given high praise to Wang Yu: a rare scholar with very high academic and business levels.
From this point of view, SenseTime is the perfect example.
For this reason, Wang Yu’s subsequent actions have always been the focus of the venture capital circle, and Wuwen Xinqiong, which is riding on the momentum of the big model and committed to solving computing power anxiety, has been in the spotlight since its establishment.
Similar to SenseTime, where Wang Yu teamed up with Yao Song, an undergraduate student from the Department of Electronic Engineering at Tsinghua University, and appointed Yao Song as CEO, this time, Wuwen Xinqiong is also a combination of famous teachers and talented students, with three Tsinghua PhDs.
First of all, Xia Lixue, the post-90s co-founder and CEO of Wuwen Xinqiong, is a PhD student supervised by Wang Yu. She received an offer from Ali Star right after graduation. She was the technical director of Alibaba Cloud's user growth products and was responsible for strategic projects such as the compression and acceleration of Alibaba Cloud's large language models and the generative AI model chip.
The emergence of the big model at the end of 2022 made Xia Lixue realize that the qualitative change of the algorithm from special-purpose to general-purpose will bring about sufficiently broad application scenarios. He then hit it off with Wang Yu, who led the Tsinghua NICS-EFC Laboratory, and became the first co-founder of Wuwen Xinqiong.
According to Tianyancha's equity penetration, Zeng Shulin, a "math nerd" from Nanning, Guangxi, is also Wang Yu's doctoral supervisor. The two indirectly hold shares in Wuwen Xinqiong through holding shares in Beijing Wuwen Qiming Enterprise Consulting Co., Ltd. Among them, Xia Lixue's shareholding ratio in Wuwen Xinqiong is estimated to be 10.9%, making her the largest shareholder.
Another chief scientist of Lianchuang, Dai Guohao, was once an assistant researcher in Wang Yu's research team. He holds a Ph.D. in engineering from Tsinghua University and holds a 6.8227% stake. He is currently a tenured associate professor at Shanghai Jiao Tong University and the head of the Artificial Intelligence Design Automation Innovation Laboratory at Qingyuan Research Institute.
With one professor and three PhDs, Wuwen Xinqiong naturally stands out in the AI VC circle that loves Tsinghua’s system.
Old shareholders are close to the water and get the moon first. Jinshajiang Venture Capital, which has tasted the sweetness of DeePhi Technology, once again took the lead in the angel round. Subsequently, more than 30 institutions divided into three groups, including national teams and local state-owned assets such as Social Security Zhongguancun Fund and Shanghai Artificial Intelligence Industry Investment Fund, Baidu, Tencent, Lenovo Capital, Xiaomi Shunwei, Softcom Hi-Tech, Zhipu AI and other strategic investors, and financial institutions such as Dachen Capital and Detong Capital entered the market.
According to ChinaVenture, more than 100 institutions contacted Wuwen Xinqiong in the last two rounds, and the subscription was oversubscribed. An insider revealed that "they were discussing business in the first half, and declining and cutting shares in the second half."
According to relevant investors' recollections, compared with pure financial investors, strategic investors (industry investors) are more likely to invest from the perspective of business empowerment at the current stage.
Building AI infrastructure, positioning itself as an "operator", and "selling" computing power to become an "invisible" unicorn
On one hand, the primary market is booming, and on the other hand, over the past 16 months, Wuwen Xinqiong has also been clearly positioning itself and accelerating its development.
The timeline goes back to the end of 2022. The rise of big models has made Xia Lixue realize that the computing anxiety faced by the AI industry has become an indisputable fact. Especially in China, the "hundred flowers blooming" at the model and chip levels in China has caused a large number of heterogeneous chips to form "ecological silos". Different hardware ecosystems are closed and incompatible with each other, which makes a lot of computing power and hardware not fully utilized.
Therefore, the original intention of establishing Wuwen Xinqiong was to solve the problem of insufficient computing power in China.
How to solve this problem? Wuwen Xinqiong chose the "light asset operation model", that is, connecting the model and hardware to build an infrastructure that makes better use of computing power.
Simply put, it is to do "computing power optimization" in the middle layer, with the goal of improving the user experience and continuously expanding the market size by optimizing usability and reducing costs.
On the technical level, Wuwen Xinqiong tackled two tough problems:
First, through the core technology of M models × N chips, different models can be quickly and efficiently deployed on various types of hardware, so as to form the best joint optimization and coordination between algorithms and computing power and maximize the value of computing power.
The second is to use a hybrid training method to enable efficient cooperation between different heterogeneous clusters, so that the efficiency and stability of use, and the effect of the cluster can reach commercial levels.
Xia Lixue said that from a technical point of view, mixed training is generally considered to be a mixture of a cluster of GPUs and another type of GPU, or another type of accelerator card. The difficulty lies in the fact that different cards have different operator libraries.
At present, Wuwen Core Qion has achieved kilocal-scale heterogeneous computing power mixed training among the "4+2" combination of six chips, including Huawei Ascend, Tianshu Zhixin, Muxi, Moore Thread, AMD, and NVIDIA. The overall efficiency of mixed training is over 95% and can reach 97.6%.
That is to say, in an ideal situation, each card is closely coordinated with each other, and a training can be completed in one month. In a non-ideal situation, Wuwen Xinqiong can complete a comprehensive large model training in less than one day and a month.
To directly connect different models and hardware, it is also necessary to predict GPU performance and break down and allocate tasks, so that the hardware can not only perform its respective functions, but also open up the communication library and coordinate well in communication.
In a cooperation case of a large model inference scenario of an Internet company, Wuwen Xinqiong "cut" 90% of the computing power used by the customer to support the same business, thereby saving computing power resources.
This is like Wuwen Xinqiong creating a "Taobao" in the field of large-model computing power. Its upstream business mainly cooperates with local intelligent computing centers that have been fully built or are planned to be built to better manage their computing power, while downstream large-model manufacturers and application parties can buy easy-to-use and efficient computing power from multiple chips with just one click.
From this point of view, Wuwen Xinqiong is moving towards an all-in-one platform, and Xia Lixue has emphasized more than once its positioning as a "computing power operator" rather than an AI infrastructure.
Wang Mengfei, SVP of strategic operations at Wuwen Xinqiong, said that AI infrastructure is more commonly understood in China as single-point optimization tools or software, while Wuwen Xinqiong's goal is to provide customers with optimized, plug-and-play computing services that are truly what they want, which is a "one-step" approach.
"Standardizing computing power sufficiently and reducing the difficulty and threshold for customers to use computing power is the difference between our vision and that of AI infrastructure companies."
Just as people don't need to know which river the water comes from before turning on the tap.
Similarly, Xia Lixue believes that in the future, when people use various AI applications, they will not know which base models are called or which accelerator card computing power is used - this is the best AI Native infrastructure.
Tsinghua University has long been competing for "computing power optimization"
Of course, although the story described by Wuwen Xinqiong is good, it is not without challenges on the market side.
In terms of business model, Wuwen Xinqiong is a complete cloud vendor model, with an IasS layer for computing power management, a PaaS layer for computing power operation and scheduling, and a MaaS layer for application vendors. The entire product matrix is relatively complete.
The company's current main source of revenue comes from the computing power cloud model, that is, the revenue from the sales of computing power. Computing power operation means integrating resources upstream, turning it into a standardized computing power product through technical capabilities, and pricing it in a unitized manner. It is then sold on the market, and the profit path is clear.
However, Wang Mengfei also admitted that the business model is somewhat similar to that of cloud vendors, and there may be some overlap or competition in the long run. For now, due to the different customer levels, positioning, and technical reserve directions, the two are more likely to cooperate than compete.
"In the past, cloud vendors transitioned from CPU to GPU, and the advantage may be that many customers can reuse them, but Wuwen Xinqiong emphasizes GPU heterogeneity and will make the technology stack more complete and in-depth, mainly to make better use of domestic computing power that is not being used well. Currently, it mainly serves customers who do large model training, and in the future, it will make technical reserves for AI application scenarios."
Compared with cloud vendors, the competition in the computing power optimization track has long been fought by Tsinghua University. In addition to Wuwen Xinqiong, industry players such as Luchen Technology, Silicon Mobility, Zhongke Jiahe, Qingcheng Jizhi, and Qingmao Intelligence are all representatives of Tsinghua University.
Source: Lieyun.com
Luchen Technology entered the market early and began to build Colossal-AI, a general deep learning system for the era of large models in 2021 to promote the application of large AI models. It is led by You Yang, a master of computer science at Tsinghua University and a Ph.D. from UC Berkeley. It has completed four rounds of financing. Investors include Sinovation Ventures, BlueRun Ventures, Sequoia China, etc.
A year later, Guan Chaoyu, who studied under Professor Zhu Wenwu, entered the company with Qingmao Intelligence to develop MLGuider - a hardware-aware automated model deployment optimization platform. The company also completed its latest round of financing in May this year.
Silicon-based Flow led by Yuan Jinhui, Wuwen Core Qion led by Wang Yu and Xia Lixue, Zhongke Jiahe led by Cui Huimin, and Qingcheng Jizhi led by Tang Xiongchao were all "involved" in 2023 and have completed a new round of financing this year.
Among them, Wuwen Xinqiong raised the most, nearly 500 million yuan, and the cumulative financing is nearly 1 billion yuan. Silicon Mobility just completed an angel+ round of nearly 100 million yuan two months ago, with a cumulative financing of 150 million yuan. It is also a core player in the primary market. More than 10 institutions including Wang Huiwen, Zhipu AI, 360, and Sinovation Ventures are its backers.
It is reported that Silicon Flow has created a standardized, ultra-high-performance generative AI Infra platform through its self-developed SiliconLLM large model inference engine and collaborative optimization of kernel, framework, mechanism and model. Its founder and CEO Yuan Jinhui once founded the Pytorch "challenger" OneFlow and invented the world's fastest large-scale topic model training system LightLDA.
Behind the Tsinghua players, the list of investors is composed of state-owned assets, industrial parties and VCs. It can be seen that the computing power business is not only to VC, but also aims at commercialization. The market here is huge, but the competition should not be underestimated.
Wang Mengfei admitted that a very important strategic layout of Wuwen Xinqiong this year is to occupy a certain market share. Conversely, it requires that a very important part of the computing power consumed by the current market must be provided by Wuwen Xinqiong as its strategic reserve.
This means that at present, whoever grabs more upstream and downstream resources first will be able to get a larger share of the "golden shovel" of computing power in the big model "gold rush".
This requires a large amount of technical investment and technical reserves, which is also where Wuwen Xinqiong will come into play after obtaining large amounts of financing.
Wang Mengfei believes that the technical side will continue to make some configurations, including various technology stacks, and talent density plays a vital role in it. It can be assumed that with the current talent density of less than 100 PhDs, Wuwen Xinqiong will continue to recruit talents.
In terms of commercialization, Wang Mengfei did not give specific figures as the company is still "young", but revealed that there is some scale revenue. Next year, the company will further expand its market share, strengthen its commercialization layout, and increase investment in ecological construction.
As Wuwen Xinqiong starts the battle for computing power within Tsinghua University, Tsinghua, standing at the forefront of the big model wave, will continue to write more chapters in the new era of AI.
Note: The cover image is from the Douban movie "Edge of Tomorrow" stills
-END-
京公网安备 11010802033920号