Arctic Xiongxin Announces Successful Delivery of Qiming 935 Series Chiplets
TapTechNews August 14th news, Arctic Xiongxin officially announced today that the Qiming 935 series of chiplets independently developed by the company have gone through nearly 2 years of design and development, and have been successfully delivered for tapeout. This tapeout has simultaneously produced two chiplets:
Qiming 935 general-purpose HUBChiplet
Ursa Major AIChiplet that natively supports all operators of the Transformer family
Arctic Xiongxin completed the test and released the first Qiming 930 chip based on the Chiplet architecture in China at the beginning of 2023, and successfully completed the process verification on the domestic packaging supply chain; in the same year, the company released the first D2D interface PBLink based on the domestic High-Speed Chiplet Interconnection Standard.
The Qiming 935 series of chiplets include a chiplet family centered around the Qiming 935 high-performance general-purpose HUBChiplet, as well as functional chiplets such as the Ursa Major AIChiplet and GPUChiplet. It is designed based on automotive-grade requirements, and can quickly generate terminal products for scenarios such as autonomous driving, embodied intelligence, and AI inference acceleration through flexible combinations. TapTechNews attaches the official introductions of the two chiplets as follows:
Qiming 935 high-performance general-purpose HUBChiplet:
It contains a multi-core ARM Cortex-A series CPU subsystem to provide core computing power, adopts interfaces and storage solutions such as PCIe 5.0 and LPDDR5, and has the ZeusGen2 NPU Core independently developed by Arctic Xiongxin for AI acceleration computing power. The Qiming 935 HUBChiplet itself can be used as an AI SoC independently.
In addition, it also includes video codec modules, ISP modules, 10 Gigabit Ethernet ports and MIPI interfaces, information security and functional safety modules, etc., which meet the design requirements of automotive grade.
Each HUBChiplet is equipped with 4 independent PBLink interfaces to provide a total of 128GB/s of chiplet interconnection bandwidth, which can be combined with other functional chiplets to expand and combine to meet the different functional requirements and computing power expansion requirements of different scenarios.
Ursa Major AIChiplet:
It contains multi-core independently developed ZeusGen2 NPU Core, and different NPUCores can either run independently or jointly run to accelerate the same task.
It natively supports all operators of the Transformer, and the ZeusGen2 NPU Core supports computing precisions such as INT4, INT8, INT16, and FP16, and supports common convolutional layers, linear layers, pooling layers, and activation layers, and functionally supports common detection and classification models including but not limited to VGG, ResNet, Yolo, etc.
It supports mainstream AI frameworks such as PyTorch, TensorFlow, PaddlePaddle, and ONNX, and the average utilization rate in different AI models has reached more than 70% in actual measurement.
In addition, the GPUChiplet applied in the field of intelligent cockpits is also being developed simultaneously, and it is expected to be delivered for tapeout in the middle of 2025.
Arctic Xiongxin introduced that the multiple self-driving chips based on the heterogeneous integration of the 935 HUBChiplet and the AIChiplet have obtained the ISO-26262 ASIL-B automotive-grade dual-certifications from SGS and China Automotive Research Institute at home and abroad in 2024, which is the first Chiplet product to obtain the automotive-grade certification in China, and it is expected to provide the core board for delivery to the downstream for adaptation work at the end of 2024.
Besides being applied in the field of intelligent driving, the AI inference acceleration module/acceleration card based on the combination of the 935 HUBChiplet and the AIChiplet can also be applied in the fields of edge-side and end-side AI inference applications (AI inference all-in-one machine, industrial control computer, robot, etc.), and can support the inference acceleration of large models of 70B.
The Qiming 935 HUBChiplet is claimed to be able to be used independently or as a general-purpose Chiplet base for expansion and development, and will become the first domestic independent general-purpose Chiplet with both CPUChiplet and MemoryChiplet functions.
