news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

author: summer

author: yoky

after large models drove the explosion of generative ai, the high demand for computing power caused chip giant nvidia's orders, revenue, and stock price to rise, while domestic chip manufacturers have been dormant.

recently, the domestic ai chips have finally been blown.

on august 26, according to the official website of the china securities regulatory commission,shanghai suiyuan technology co., ltd. (hereinafter referred to as "suiyuan technology") has completed the listing guidance filing with the shanghai securities regulatory bureau, and with cicc as the guidance institution, it has started the a-share ipo journey.

if suiyuan technology is successfully listed, it will become the "second ai chip stock" after cambrian's listing in 2020.

according to public information, since its establishment in march 2018, suiyuan technology mainly provides ai training and reasoning products, covering chips, boards, intelligent computing all-in-one machines, liquid-cooled computing clusters and supporting software systems. so far, it has raised 10 rounds of financing, totaling 7 billion yuan, and its valuation has reached 16 billion yuan. investors include the national big fund, tencent, meitu, zhen fund, redpoint china, etc. among them, tencent has invested in 6 rounds in a row and has become the largest shareholder of suiyuan technology, holding more than 20% of the shares.

both founders of suiyuan technology are "veterans" in the semiconductor industry. before establishing suiyuan, zhao lidong and zhang yalin both had more than 10 years of work experience at amd and successively held senior positions. after zhao lidong returned to china, he was in charge of semiconductor investment-related work at unigroup communications technology, accumulating the domestic upstream supply chain for the establishment of suiyuan.

it is worth noting that suiyuan technology has received high attention from the entire industry in the early stages of ipo guidance.whether china's big models can break out of the independent curve depends to a certain extent on the performance breakthrough and evolution of the underlying chips. this has also directly led to the fact that every move of every ai chip company has been put under the spotlight.

from the perspective of industry development, the massive computing demand during the pre-training phase of large models is only the first wave, and there will be commercial deployment, iterative training, industrial applications, etc. therefore, zhang yalin believes that "the demand for computing power will become greater and greater in the future, far beyond what we can imagine now."

however, the market situation of the entire semiconductor and chip industry is not so optimistic. after the explosion of the number of companies in the short term, there will be a period of elimination, and finally convergence to a few leading manufacturers. it can be said that the tickets to enter the market are scarce. zhao lidong also publicly said: "the characteristic of the chip industry is that the boss makes money, the second can also make money, and the third is very difficult, and basically there is no market."

can suiyuan, which was the first to go public, take the second place?

1

a chip company that only does "ai"

currently, the semiconductor industry has been shrouded in a fog in the capital market.

since june 2023, the phased tightening of a-share ipo policies and the cyclical downturn in the global semiconductor industry have caused domestic chips to be cold again after being popular in the capital market. according to incomplete statistics from public data, as of july 29, 43 semiconductor companies have terminated their ipos in 2024, and 42 have voluntarily withdrawn their ipos, including 19 chip design companies.

among ai chips, apart from cambrian, which successfully landed on the science and technology innovation board in 2020, there was no news about biren after it was rumored to be planning to be listed on the hong kong stock market in july last year. at this time, suiyuan's ipo, to some extent, is a boost to the confidence of the entire industry.

one of the core reasons why suiyuan technology was able to impact the a-share market in an obstructed environment is that the explosion of generative ai has created a demand gap for high-performance ai chips, and suiyuan happened to be betting on cloud ai chips from day one.

zhao lidong once shared his entrepreneurial thoughts: when he started his business in 2018, he had several options in front of him, including cpu, gpu and ai chips. "at that time, the cpu and gpu solutions were already very mature, and i wanted to make ai chips."

in simple terms, nvidia's almost monopolized gpu is more like a "jack of all trades", while ai chips are optimized for specific algorithms such as machine learning and deep learning. in this scenario, they are better than gpus and consume less power, but they are not universal. usually, companies use a combination of gpu + ai chips to reduce the cost and high power consumption of gpus during training.

focusing on ai training and reasoning, enflame technology launched two generations of the suisi series chips, suisi 1.0 and suisi 2.0, from 2018 to 2021. on this basis, it launched the cloudblazer ai accelerator cards cloudblazert20, i10, t21, i20, and has achieved mass production and large-scale implementation.

hardware products include boards, intelligent computing all-in-one machines, and liquid-cooled computing clusters. the hardware is also equipped with two software development platforms, yusuan and jiansuan, to provide an integrated solution of self-developed chips + accelerators + software systems.

high energy efficiency and low power consumption are the differences emphasized by suisi chips.the suisi chip uses the self-developed sia (syntiant neural decision processor) architecture, which is a low-power ai processor that uses an event-driven approach to ensure a balance between low power consumption and high performance on the basis of high precision. at the same time, the sia architecture has built-in multiple sensors, including sound, image, etc., which can realize the perception and recognition of multiple ai scenarios.

2022 is a key node in the development of suiyuan technology. under the macro policy of the country's release of the east-west computing strategy, suiyuan technology has begun to launch the yunsui intelligent computing cluster, which is mainly aimed at large-scale and intensive artificial intelligence computing power application scenarios, such as smart cities, smart transportation, and intelligent computing centers.

this has enabled the accumulation of the previous two generations of chips and accelerators to begin to be put into large-scale use. public information shows that suiyuan technology has assisted various cities in gansu, an important province in eastern and western computing, including qingyang and zhangye, to build ai training computing power clusters of thousands or tens of thousands of calories.

after 2023, the explosion of generative ai has put forward higher requirements on computing power and speed. in order to more intuitively demonstrate the systematic solutions and chip capabilities of suiyuan technology, suiyuan technology launched the wenshengtu maas platform service product suiyuan yaotu, using suiyuan technology's self-developed suisi series chips as computing power support.

zhang yalin revealed that suiyuan technology will soon release a platform called wenshengwen. "the purpose of this platform is to allow suiyuan's systematization and solution capabilities to directly reach high-level users through the display of this product."

from the underlying chips to the boards, the integrated hardware and software solutions, and the final display of effects, the reason why it is so "big and complete" is essentially limited by the entire chinese chip environment."if you are positioning yourself as a chip company or a hardware provider, it is difficult for you to communicate with the application and model parties in a concrete way. however, the construction of such a systematic product allows you to easily reach more customers in the form of product prototypes," said zhang yalin.

1

tencent's chip is the key

it is noteworthy that after six rounds of investment, tencent became the largest shareholder of enflame. for tencent, which started late in developing its own chips, enflame is more like a layer of insurance for itself.

among the major internet companies, chip layout has become a consensus among the giants. after alibaba acquired zhongtianwei, it reorganized internally into pingtou ge. baidu owns kunlun core, and huawei has even built a fully self-developed ascend ecosystem.

tencent has placed its bets on suiyuan technology for the lifeline of ai chips.as early as the prea round, tencent joined the shareholder camp of suiyuan technology and has never missed any round of financing since then.

in addition to being a major shareholder, tencent is also a major customer and business partner of suiyuan technology, contributing both money and manpower.

in the cooperation between the two parties, tencent's huge business system has provided suiyuan technology with a wealth of chip usage scenarios and training opportunities. for example, in the human-machine battle link, with the help of the yunsui artificial intelligence training acceleration card as the computing power base, by accelerating the training of cloud ai models, machines can continuously learn game dynamics and diverse battle roles and strategies at a faster speed. in addition, tencent's speech-to-text, tencent conference and other products also use "suisi" chips on a large scale.

in addition, suiyuan has also gone to the cloud. in 2023, tencent cloud and suiyuan technology jointly created a hybrid cloud architecture with "storage and computing separation". while storing the code locally, cloud computing power is embedded in the entire design and verification process, efficiently supporting the improvement of r&d efficiency, helping suiyuan save tens of millions of it investment, increase task concurrency by 100%, and shorten the overall simulation cycle by 30%.

regarding the cooperation with tencent, zhang yalin once mentioned that "full matching deployment with large internet strategic customers is a very critical handle for us in reasoning and deployment. because large internet company strategic customers have all the real content understanding, intelligent interaction, intelligent conferencing, recommendation search, large model applications, including games and other scenarios. they will pay great attention to cost performance and require you to achieve better cost performance than your competitors before they will conduct a trial."

it can be said that tencent has provided a huge "test field" for suiyuan, and for tencent, what is more important is to solve the shortcomings of self-developed chips through investment. in 2021, tencent cooperated with suiyuan technology to develop a zixiao series chip, which has achieved more than twice the cost performance of similar products in the industry in terms of ocr text recognition, smart conferencing, image and voice noise reduction, etc., and through the zixiao chip, tencent can provide inclusive and high-quality artificial intelligence solutions to customers in the financial, transportation and other industries.

today, as the competition for big models intensifies, it has entered the battle for infrastructure.the weakness of self-developed chips means that it may be "choked" by other companies at any time. this is also the key reason why tencent needs a strong bond with suiyuan technology.

1

internet on the left hand, intelligent computing center on the right hand

the chip industry is a long-cycle, high-investment industry and it is difficult to achieve economic benefits in the short term, so commercialization has always been the biggest challenge.

according to the official website, suiyuan technology's main customers are divided into four parts: internet, smart city, intelligent computing center, and finance. the core customers can be divided into two categories: one is internet companies represented by tencent and meitu; the other is government projects such as smart cities and intelligent computing centers jointly built with local governments.

two of the largest customers, tencent and meitu, are core shareholders of suiyuan. with the help of the shareholders' business scenarios, suiyuan technology has been able to polish its own products and accumulate service experience from r&d, design to deployment and application.

in 2018, when accepting tencent's pre a round of financing, zhao lidong mentioned, "in order to achieve a breakthrough in business implementation, we must seek cooperation. only in this way can we make the first step of implementation successful." in 2019, suiyuan technology launched its first generation of chips. compared with creating a finished chip and then looking for customers, suiyuan technology cooperated with tencent to shorten the research and development of a generation of chips by several months.

for chips, business scenarios are very important."chips become more useful the more you use them."zhao lidong said, "the products of international mainstream chip manufacturers have gone through more than ten years or even decades of iterative optimization to reach their current level and scale."

in addition, some large government projects have also made use of the cooperative resources of the listed company hongxin electronics. public information shows that the two have in-depth cooperation and have jointly created a domestic computing power platform suitable for artificial intelligence applications and basic training capabilities, as well as the gansu east and west computing power network.

with a large internet company on the left hand and an intelligent computing center on the right hand, suiyuan has had stable income and scenarios in the past six years, and also laid the foundation for its listing at this time.

after entering 2023, the emergence of a large number of new chinese base model startups will also provide new market opportunities for underlying ai chip companies.

during the 2024 waic, zhang yalin introduced the current customer structure of suiyuan technology: "at present, in china, first of all, most of the users with large computing power consumption needs are still in the internet; secondly, they may be vertical industries and central state-owned enterprises, and finally, start-up large model companies, etc."

in 2024, zhipu ai codegeex and suiyuan technology yunsui i2.0 tried to cooperate. the "suiyuan + zhipu ai" ai large model programming assistant all-in-one machine will help developers improve programming efficiency, shorten the development cycle, and provide intelligent programming assistance functions. however, this has not entered the core training and reasoning stage of zhipu ai, but only some point-like attempts on the business side.

however, as large models enter the application, more and more emerging companies are also challenging the service capabilities of chip manufacturers. according to media reports, a domestic ai chip company that does not want to be named revealed that since 2023, it has tried to let backstage r&d personnel go to the front line to provide personal services - not only to be on-site for joint training, but also to set up a small group with r&d personnel for customers after sales orders. customers with a unit price ranging from several million to hundreds of thousands can enjoy 24/7 consultation.

if we say that chip manufacturers were previously in the 1.0 stage of volume technology and volume products, then after 2024, they will enter the 2.0 stage of volume market and volume services.

going public is just the beginning of the next stage.