withBig dataThe time has come,Artificial intelligenceTechnology is gradually entering various fields, and the demand for computing power is becoming higher and higher. However, in the process of large model training, due to the huge amount of data required, the demand for computing power has also increased, which has become an urgent problem to be solved. Although the country is in AIChipsTremendous progress has been made in the development and production of the company, butDomestic chipsIt has not been widely recognized by the market. Therefore, the experts at the meeting called for domestic AI to solve this computing power problemChipsThe system needs to reshape the computing power ecology of large models to improve their competitiveness.
For nowDomestic chipsIt is relevant in large model trainingSoftwareThe porting is not smooth, and experts believe that the underlying research and development, including programming, should be further strengthenedframe, parallel acceleration, communication libraries, etcSoftwareR&D work to enhance domestic AIChipsof "inclusiveness". Professor Zheng Weimin said that ifDomestic chipsThe system is able to be easily transplanted based on foreign countriesChipsofSoftware, the user will feel itDomestic chipssuperiority.
In order to develop AIChipsSome large international enterprises can even abandon some high-profit projects and programs to promote the innovation power of the entire industry. For example, international businesses can take the initiative to open source sectionsSoftwareThe bottom **, let the profit be based onChipsofSoftwaredevelopers, thus forming with AIChipsas the core computing power system.
In order to improve the overall computing powerEfficiencyTo meet the differentiated requirements, it is recommended to establish an "intelligent computing center" to support the calculation of trillion-level parameters of large models to achieve large-scale deployment. The center can make use of the existing high-quality data center to meet the needs by improving networking capabilities, collaboration capabilities, and software and hardware adaptation capabilities. At present, some cities are developing "Wanka" level intelligent computing centers.
Experts at the meeting believed that promoting the integration and innovation of existing computing power will further improve the use of computing powerEfficiency。In connection support andParallel computingand other aspects, the most efficient way is needed to support multiple requirements such as model parallelism and data parallelism. At the same time, there are more complex networks involved behind the large modelArchitectureand intelligenceAlgorithmsWe need to lower the threshold for the use of large model computing power and improve the overall situationEfficiency
In order to solve the current problem of tight computing power of large models, an effective way is to adopt the "cloud" model. By sharing and integrating the existing computing power, the Vanka platform and the large model training and inference platform equipped with an efficient network interaction mechanism can be opened to large enterprises, small and micro enterprises for common use. Build a basic ecosystem such as a cloud platform, and migrate and train data through riggingDatabasesand other tools, so that computing power can be used as much as water and electricity.
In the cloud mode, by breaking through the asymmetry between supply and demand, the inclusiveness of computing power is realized, so that more enterprises and institutions can benefit from powerful large-scale computing power. This will promote the development and innovation of the industry and accelerateArtificial intelligenceTechnicalApp promotion
By reshaping the computing power ecology of large models, strengthening the underlying research and development, promoting the integration and innovation of existing computing power, and breaking through the asymmetry between supply and demand, we will promote the inclusiveness of computing power and domestic AIChipsThe system will be able to better solve the current problem of computing power shortage and enhance its competitiveness. At the same time, it is also a new opportunity, providing space for domestic enterprises and institutions to develop and innovate. We have reason to believe that by working together, China's AIChipsThe system will practice "internal strength", inArtificial intelligenceMore brilliant achievements have been made in the field.