after 16 months of establishment, it has received 1 billion yuan in financing. wuwen xinqiong aims to become the preferred "computing power operator" in the era of large models
2024-09-02
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
photo courtesy of this newspaper
infinigence ai, a startup company that provides large-scale model infrastructure, announced today that it has completed nearly rmb 500 million in series a financing. so far, infinigence ai, which has only been established for 16 months, has raised nearly rmb 1 billion in financing.
xia lixue, co-founder and ceo of wuwen xinqiong, believes that the new "80/20 rule" brought about by the ai 2.0 wave only requires solving 20% of key technical problems to support the generalization of 80% of vertical scenarios. china's computing power ecology is facing the current situation of unbalanced supply and demand and uneven distribution of resources. this undoubtedly creates an era opportunity for wuwen xinqiong to drive upstream and downstream cooperation to achieve efficient integration of diversified heterogeneous computing power.
wuwen xinqiong said that the funds raised in this round of financing will be used to strengthen the recruitment of technical talents and technology research and development, maintain the technological leadership of software and hardware collaboration and diversified heterogeneity; deeply promote the commercial development of products, and maintain the close integration between infini-ai heterogeneous cloud platform products and the market; strengthen ecological cooperation, activate heterogeneous cluster computing resources, build an ai computing power base that supports "m types of models" and "n types of chips", and serve as a "super amplifier" for ai model computing power... wuwen xinqiong will be committed to becoming the preferred "computing power operator" in the era of big models.
become a "super amplifier" of ai model computing power
the actual industrial scale that a large model can support depends on the actual available computing power of the ai model. wuwen xinqiong believes that the actual available computing power of a large model depends not only on the theoretical computing power of the chip, but also on the efficiency of computing power utilization through the optimization coefficient, and the overall computing power scale through the cluster scale. therefore, wuwen xinqiong proposed the formula of "chip computing power × optimization coefficient (software and hardware collaboration) × cluster scale (multiple heterogeneous) = ai model computing power". following this formula, wuwen xinqiong will continue to improve the utilization rate of chip computing power in large model tasks through software and hardware joint optimization technology, and improve the utilization rate of cluster computing power through multi-heterogeneous computing power adaptation technology, thereby expanding the overall computing power supply of the industry.
in terms of joint optimization of software and hardware, wuwen xinqiong has greatly improved the utilization rate of mainstream hardware and heterogeneous hardware through its self-developed reasoning acceleration technology flashdecoding++, surpassing the previous sota, and completed the adaptation of multiple mainstream open source large models on more than 10 computing cards such as amd, huawei ascend, biren, cambrian, suiyuan, haiguang, tianshu zhixin, muxi, moore thread, nvidia, etc., and achieved the industry's first reasoning acceleration results on some computing cards, efficiently meeting the increasingly high large model reasoning needs of various industries. based on the optimization effect achieved by this solution, wuwen xinqiong has signed a strategic cooperation with amd to jointly promote the performance improvement of commercial ai applications.
in terms of adapting multiple heterogeneous computing power, wuwen xinqiong also possesses rare heterogeneous adaptation and cluster capabilities in the industry. the large-scale heterogeneous distributed hybrid training system hethub released in july is the first in the industry to achieve kilocal-scale heterogeneous computing power hybrid training between the "4+2" combination of six chips including huawei ascend, tianshu zhixin, muxi, moore thread, amd, and nvidia. the cluster computing power utilization rate reached as high as 97.6%, which is about 30% higher than the benchmark solution on average. this means that under the same multiple chip computer room conditions or cluster conditions, wuwen xinqiong can compress the total training time by 30%.
providing full-stack capabilities from heterogeneous computing power to large-model application development
in recent years, the international model layer and chip layer have gradually formed a "double-headed convergence" pattern, while china's model layer and chip layer continue to present an "m×n" pattern consisting of "m types of models" and "n types of chips". however, different hardware platforms need to adapt to different software stacks and tool chains, and there has long been an "ecological silo" phenomenon between heterogeneous chips that is difficult to use. as more and more domestically produced heterogeneous computing power chips are used in local computing power clusters across the country, the problem of heterogeneous computing power being difficult to effectively utilize is becoming increasingly severe, and has gradually become a bottleneck for the development of china's large model industry.
relying on the advantages of software and hardware collaboration and diversified heterogeneous technology, wuwen xinqiong has built the infini-ai heterogeneous cloud platform based on a diversified chip computing power base. the platform is backward compatible with diversified heterogeneous computing power chips, which can effectively activate dormant heterogeneous computing power across the country. the computing power currently in operation covers 15 cities across the country. in addition, the infini-ai heterogeneous cloud platform also includes a one-stop ai platform (aistudio) and a large model service platform (genstudio). since the platform was launched, many leading customers in the large model industry, such as kimi, liblibai, liepin, shengshu technology, and zhipu ai, have been using heterogeneous computing power on the infini-ai heterogeneous cloud platform stably, and enjoying the large model development tool chain service provided by wuwen xinqiong.
infini-ai heterogeneous cloud platform can not only help downstream customers easily shield hardware differences and use the powerful capabilities of underlying heterogeneous computing power efficiently and seamlessly, but will also effectively break the ecological dilemma of domestic heterogeneous computing power, accelerate the gradual migration of upper-level applications to heterogeneous computing power bases, effectively integrate and expand the scale of available computing power for domestic large model industries, and truly transform heterogeneous computing power into usable, sufficient, and easy-to-use large computing power, helping to build a localized heterogeneous computing power ecosystem with chinese characteristics.
author: shen qiusha
text: shen qiusha photos: provided by the interviewee editor: shen qiusha responsible editor: ren quan
please indicate the source when reprinting this article.
