According to news from Gamingdeputy on February 20th, according to the news from Arctic Xiongxin Information Technology Co., Ltd., recently, Beibei Xiongxin released simultaneously at the Xi’an Qinchuangyuan Artificial Intelligence Cutting-edge Scientific and Technological Achievement Conference and the opening ceremony of Beijing Weihao Chuangxin Incubator.The first “Qiming 930” chip based on Chiplet architecture.
According to reports, the central control core of the chip adopts RISC-V CPU coresAt the same time, it can be equipped with multiple functional core particles through the high-speed interface. Based on the substrate material produced in China and 2.5D packaging, the computing power can be expanded, and it can be used in different scenarios such as AI reasoning, privacy computing, and industrial intelligence. At present, it has cooperated with many AI downstream scenario partners for testing.
IT home popular science, Chiplet architecture refers to theSplit large chips into small chips for production and integrated packagingwhich can effectively improve the comprehensive yield rate of high-power chip manufacturing, and create flexible collocation options through chip reuse. At present, Intel and AMD’s products use related technologies, which are improvements to traditional single-chip solutions.
Arctic Xiongxin stated that the company’s Chiplet solution decouples the general requirements of downstream scenarios from the special requirements, and designs and manufactures small chips separately and integrates them, effectively solving the problems of downstream customers in algorithm adaptation, iteration cycle, computing power utilization, and computing power. The core pain point that is difficult to balance in various aspects such as cost.
Qiming 930 uses 12nm process to produce, HUB Chiplet adopts RISC-V CPU core, which can provide 8~20TOPS (INT8) dense computing power by flexibly carrying multiple NPU Side Dies. Qiming 930 can be used independently as an AI accelerator card, and can also be integrated with a variety of functional Side Dies through D2D expansion, and has a variety of product forms.
The Qiming 930 chip uses the self-developedThe third generation “Moos” core architecture NPUsupports INT4, INT8 and INT16 calculation precision, supports common convolutional layers, linear layers, pooling layers and activation layers, functionally supports common detection and classification models, including but not limited to VGG, ResNet, Yolo, etc.

Different NPU cores can run independently or jointly to accelerate the same task, which is flexible and convenient to use. In the design stage of compilation and deep coupling of the architecture, and in-depth consideration of the constraints and costs caused by the multi-chip interconnection, after optimization, the utilization rate of chip resources is claimed to be up to 70% on average.
packaging,Qiming 930 chip adopts domestic substrate and 2.5D packageand optimize the interface and packaging scheme according to the characteristics of domestic substrates to ensure the transmission efficiency of multi-chip interconnection.
In addition, the Qiming 930 chip supporting package includesBasic driver, framework support, application layer full-stack application deployment toolto provide customers with a debugging-deploying-application-oriented software stack, and has successfully verified a series of models with partners.
