news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

(image source: hrexchangenetwork)

a domestic gpu (graphics processing unit) manufacturer that claimed to be the "nvidia" of china collapsed overnight.

titanium media app reported on august 31 that according to multiple industry sources,xiangdixian computing technology (chongqing) co., ltd. (hereinafter referred to as "xiangdixian"), a gpu company that has raised 2.5 billion yuan and was once valued at 15 billion yuan, held a full-staff meeting on august 30, announcing the official dissolution of the company and the termination of labor contracts for all employees.

this is the latest company to go bankrupt and dissolve among a large number of domestic gpu startups that have emerged recently.

according to titanium media app, the management of xiangdixian revealed at a staff meeting that the company signed a b-round bet agreement of 500 million yuan with the investor, but ultimately failed to reach the terms of the agreement, so it was sued by shareholders. currently, the company's account has been frozen, and there is no incoming funds now or in the future. therefore, all employees have been terminated and unpaid wages are recorded in the account. the company will continue to seek new investors and sue shareholders to negotiate the account freeze and return all equity funds to employees, but the company did not disclose the specific implementation plan, nor did it disclose more details of the bet agreement.

information from maimai and xiaohongshu shows that many employees said that xiangdixian still owed them two months' salary and the company would not provide compensation.

there are also reports that the layoffs involve all 400-plus employees. titanium media app according to the company's recruitment presentation at the university of electronic science and technology in 2021, the company's size has exceeded 200 people, of which one-third of the more than 60 employees in chengdu are master's graduates from the university of electronic science and technology. however, industrial and commercial information shows that in the 2023 annual report, the number of insured persons in xiangdixian chongqing entity was only 84, and the previous highest was 255. other subsidiaries and affiliated corporate entities have not disclosed specific insurance information. it is estimated that the number of layoffs this time may be hundreds.

on the morning of august 31, the editor of titanium media app sought more information from xiang dixian, but no response was received as of press time.

less than four years after its establishment, xiangdi went bankrupt before its two chips were commercialized on a large scale

public information shows that xiangdixian was established on september 29, 2020 with a registered capital of approximately 17.83 million yuan. it is headquartered in chongqing and has established r&d centers in beijing, shanghai, chongqing, chengdu, suzhou and other places. it is a design company that develops cpu (general-purpose processor)/gpu chips. it mainly develops high-performance, low-power processors and related special chip products with completely independent intellectual property rights for desktops, workstations, edge computing and other fields. the company has also received support from the chongqing municipal government. it is a chip science and technology innovation enterprise that chongqing has focused on incubating. the latest valuation has reached 15 billion yuan, making it one of the few unicorns in chongqing's integrated circuit field.

tang zhimin, chairman of xiangdixian company, is currently also a doctoral supervisor at the institute of computing technology, chinese academy of sciences. he holds a ph.d. in engineering from the institute of computing technology, chinese academy of sciences, and a bachelor of science in computer science from nanjing university. he has served as the head of the loongson project, general manager and chief scientist of haiguang information, and cto of sugon. he is the leader of the loongson research group established by the chinese academy of sciences in 2001. he has led domestic high-end general-purpose chip projects such as the loongson 1 and 2 cpus, haiguang series cpus, and haiguang dcu, and achieved success.

in 2020, tang zhimin founded xiangdixian and switched to gpu products.

at the product level,in 2022, based on the ip core license of imagination technologies in the uk, xiangdixian's second gpu chip project (pangu a0) was successfully taped out and released in chongqing, named "tianjun no. 1". xiangdixian said that the "tianjun no. 1" gpu is based on the pangu microarchitecture, equipped with 2048 computing cores, fp32 single-precision floating-point computing power of 4tflops, and ai computing power of 16tops. it is a domestically produced high-performance general-purpose graphics processor with completely independent intellectual property rights, and its performance has reached the level of nvidia gtx1660.

tang zhimin said at the time that gpu chips are important components widely used in desktops and data centers. after more than 300 r&d personnel spent 15 months on research and development, the tianjun-1 gpu chip will be mass-produced, thus filling the gap in the domestic high-performance general-purpose gpu chip market and facilitating the informatization and intelligentization of industries such as finance, energy, transportation, and medical care. at the same time, the pangu architecture used by the tianjun-1 gpu chip has accumulated nearly 100 technical invention patents, laying a solid foundation for the development of subsequent new products.

in september 2023, xiangdixian's "tianjun-2" image gpu chip was officially released. it is positioned as a cost-effective, low-power, autonomous and controllable gpu. in terms of product design, it adopts the pangu microarchitecture, up to 1024 computing cores, and 64-bit gddr6 ultra-high frequency specifications. the maximum video memory capacity can reach 8gb and the bandwidth can reach up to 128gb/s.

however, at present, xiangdi’s two gpu products, "tianjun-1" and "tianjun-2", have not yet achieved large-scale mass production and commercialization. its public partners include h3c, tsinghua tongfang, powerleader, sumifangtian technology, kylin, tongxin, china science and technology fangde, aerospace information, pudong development bank, national new energy vehicle technology innovation center, etc.

in terms of financing,industrial and commercial information shows that in september 2020 and august 2022, xiangdixian completed the company's angel round and a round of financing respectively. the amount of financing was not announced. investors included; in july 2023, xiangdixian received investment from citic group, qianzhan investment, dianbao fund, yarui capital, founder hesheng, langkong asiana, and liangjiang capital; in october 2023, the company received investment from jiangsu sino-german service trade industry investment fund (limited partnership); in december 2023, it received investment from yuanxu fund, shengshi investment, qianshan capital, shuimu chunjin capital, qianhai yangtze river fund, and boyu private equity.

according to statistics from titanium media app, the number of shareholders under the entity of xiangdixian computing technology (chongqing) co., ltd. reached 22.

information on the list of chongqing's unicorn companies in 2024 shows that xiangdixian company is valued at us$2.169 billion (approximately rmb 15 billion).

in fact, there were early signs of the disbandment of all members of xiangdixian.

in april this year, the market reported that xiangdixian laid off all fresh graduates, and in may, there was news of "salary cuts for all employees". according to laid-off employees, the company claimed that layoff compensation would not be given until september 30th, but now there is still a month to go before that date, but news of the company's dissolution has been reported. just one day before the all-staff meeting on august 29th, multiple messages on maimai said that "all xiangdixian employees have graduated". and many laid-off employees said that they were pessimistic about the severance pay.

as of press time, xiangdixian was involved in a labor arbitration-related case filing information, with the case number (2024) chuan 0191 minchu 14123, which is currently under trial by the people's court of chengdu hi-tech industrial development zone.

some industry insiders believe that not only did xiangdixian "fail" the bet with the lead investor, but there were also certain problems in the company's internal governance, changes in the external environment, and the lack of a large amount of commercial revenue. due to the company's deficiencies in marketing and commercial operations, its self-generating ability is limited. due to the external environment, it is difficult to raise funds to survive, and it did not get more funds in time for research and development and rapid commercialization, which ultimately led to the company entering a "dormant" and dissolution state. it was actually expected that it would take the last step.

the continued recovery of chips may face challenges, and the development of domestic semiconductors has a long way to go

the disbandment of xiangdixian is not an isolated case.

on the evening of august 19, the listed company dongxin semiconductor issued an announcement stating that the company plans to increase its capital in domestic gpu manufacturer lisuan technology (shanghai) co., ltd. (hereinafter referred to as "lisuan technology") through its own funds of 200 million yuan, and subscribe to its newly added registered capital of 5 million yuan. after this capital increase, dongxin semiconductor will hold approximately 37.88% of the shares of shanghai lisuan.

in this round, in addition to dongxin semiconductor, other investors jointly increased capital in shanghai lisuan with the same pre-investment valuation. other investors plan to increase capital in shanghai lisuan by a total of 128 million yuan and subscribe to its newly increased registered capital totaling 3.2 million yuan.in other words, lisuan technology received a total of 328 million yuan in financing this time.

earlier, lisuan technology encountered problems with poor financing and was exposed to have "burned" 300 million yuan in financing and was on the verge of bankruptcy. many r&d employees claimed on social platforms that they were laid off, involving nearly 100 employees.

several industry insiders familiar with the above transaction told titanium media app that after the completion of this investment, dongxin will become the largest equity/controlling shareholder of shanghai lisuan. lisuan technology's previous financing was spent due to multiple factors such as r&d obstruction during the 2022 epidemic, the semiconductor industry downturn, and the us department of commerce's export control on gpus. at present, the first gpu chip developed by lisuan technology has been taped out, and it is expected to be put into mass production after dongxin joins.

in addition to lisuan technology, the internal situation of other domestic gpu and general gpu companies is not as good as expected, mainly due to the impact of the us department of commerce's semiconductor export control. tsmc is unable to manufacture advanced process orders for companies, and gpus themselves require high-precision and cutting-edge processes. in addition, many of them are customized from other ip cores, and the accumulation of core technologies is weak. titanium media app learned that some gpu chip companies have shifted their r&d investment to unrestricted areas such as ai chip software stacks, algorithms, and components.

according to ee times, imagination ceo simon beresford-wylie once said: "the chinese market is more challenging than we thought because the uk's export control system and china's entity list have an impact on us and our revenue." it is reported that imagination is a british chip design company controlled by chinese capital. more than 80% of domestic gpu companies, including xiangdixian, use imagination img b/c's ip core license for customized development.

"many domestic companies are using gpu ip licensing, but this type of chip is not industry-independent, and overseas ip is also difficult to guarantee the information security of domestic customers."lv jianping, cto of gpu company tianshu zhixin, once told titanium media app that the chinese gpu track cannot ignore the importance of "full independent development and wide application of technology."

lv jianping frankly said that gpu in a broad sense has three major challenges: first, the hardware foundation of graphics, second, the scalability required for scenarios such as (from mobile phones and desktops to) data centers, and third, general computing. all three technologies are difficult. whether it is gpu, dpu, or ipu chips, they are all high-tech fields, and it is difficult to simply describe them as easy or not.

"semiconductors must be an industrial system with long-term technology accumulation and a long harvest period. we should take the right path and not be led astray by capital or internet models," lu jianping emphasized.

at present, china's chip semiconductor industry as a whole is showing a recovery trend, but as the industrial environment changes, this recovery is also becoming more complex.

statistics released by cinno research show that from january to june 2024, the investment amount of china's semiconductor projects was approximately 517.3 billion yuan (including taiwan, china, the same below), a year-on-year decrease of 37.5%, and the scale of industry investment has declined.

among the main sub-categories of capital flows, investment in semiconductor materials saw the largest decline, with an investment amount of rmb 66.81 billion, accounting for approximately 12.6%, a year-on-year decrease of 55.8%; followed by investment in wafer manufacturing, with an amount of approximately rmb 246.8 billion, accounting for approximately 47.7%, a year-on-year decrease of 33.9%; in addition, chip design and packaging and testing declined by 29.8% and 28.2%, respectively.

titanium media app once exclusively counted that in 2023, 10,900 chip-related companies in china had their industrial and commercial registrations cancelled or revoked, a year-on-year increase of 69.8% and an increase of 89.7% from 5,746 in 2022. according to calculations, on average, more than 31 chip companies have their industrial and commercial information cancelled or revoked every day. according to the editor of titanium media app, in the past five years, the number of chip-related companies in china that have been revoked or cancelled has exceeded 22,000.

behind the continuous tightening of capital and scale, whether the industry is returning to healthy development or lowering market expectations due to weak recovery has become the focus of a new round of attention from all parties.

according to china business news, he hui, director of semiconductor industry research at omdia, pointed out that the decline in industry investment is not contradictory to recovery, and that it is normal for capital to decline in scale as the domestic semiconductor development has entered a new stage of integration in recent years. in addition, since last year, the country has released some relatively obvious signals in terms of policies and actions, announcing that the semiconductor industry is entering a period of integration.

in the long run, china's ai chips still have great market opportunities.

according to statistics from titanium media app, the global ai chip market will be worth about $53.6 billion in 2023, with the combined market size of ai chips in china and the united states accounting for more than 90% of the global total. among them, china's ai chip sales are about 120 billion yuan (about $16.54 billion), and the united states accounts for more than $35 billion.gartner predicts that by 2027, global ai chip revenue will more than double that of 2023, reaching $119.4 billion.

gartner predicts that generative ai will be used by more and more people starting from 2024, and the demand for cloud ai inference chips will exceed that for training chips in 2025, and the demand will increase faster than training. especially compared with the large investment in inference costs of cloud computing, from the perspective of different model parameters, edge/end ai is more feasible. the device side can support models with a scale of 1 billion to 10 billion parameters, and the edge side can support large models with a scale of 10 billion to 100 billion parameters.

roger sheng, vice president of gartner research, told titanium media app earlier that it is meaningless for some companies to keep a low profile and not dare to do business with the country. as long as large-scale applications are adopted and product competitiveness is improved, the united states will definitely know that it is not because the company has not developed ai chips, but because the company has not posed a threat to it. some companies need to adjust their cognitive biases. as long as the chips are well made, the domestic market will give them opportunities.the united states is indeed relatively ahead in the field of ai, but when it comes to basic productivity, chinese companies have great capabilities to realize these products.

xinhua news agency once pointed out thatif technology companies want to catch up in the big model competition, they must build powerful computing facilities, and ai chips are becoming a bottleneck. it is estimated that nvidia's ai chips currently account for 70% to 80% of global sales in this field.

regarding the us semiconductor export control, the global times commented on august 31 that the us has taken more targeted measures to restrict china's development of the semiconductor industry, which may bring new challenges to china's high-tech industry, but at the same time it will also inspire the determination and confidence of chinese companies and scientific research institutions to tackle core technologies. the us, in the name of security, willfully cuts up the global supply chain, which is bound to hurt the interests of western companies themselves.

"china's ai big model is from 'nothing' to 'independent research and development'. ai chips also have this firm confidence. only then can we build the ai ​​chip industry," said sheng linghai.