央视画面披露，阿里旗下平头哥PPU芯片在部分重要参数上比肩英伟达的H20芯片，并超过H800芯片。 H800和H20均基于英伟达H100改版而来，是为了满足美国出口管制而推出的中国大陆市场"特供"版芯 片。 平头哥PPU集成HBM2e（第三代高性能内存），和H800相同，但落后于H20的HBM3（第四代高性能内 存）。但PPU的显存容量与H20相同，均为96G。卡间互联带宽上，PPU为700GB/s，介于A800和H20之 间。PCIe（外围组件快速互连）接口层面，PPU也优于A800，与H20等同。功耗上，PPU与A800一致， 均为400W，落后于H20的550W。 央视《新闻联播》画面。 如果和国产阵营的华为昇腾910B芯片对比，央视画面显示，平头哥PPU在上述所有性能参数指标上均处 于领先。不过，昇腾系列最新款是910C芯片。 市场传闻半个月之久的阿里自研AI芯片，9月16日晚出现在央视《新闻联播》。 有媒体此前8月下旬报道，阿里已开发出一种比其旧款芯片更通用的新款芯片，用于AI推理任务，以填 补英伟达留下的市场空白。 9月11日，硅谷科技媒体The Information进一步爆料称，阿里和百度都在 ...

Core Viewpoint - Shanghai Jiyue Xingchen Intelligent Technology Co., Ltd. has launched its first full-size, native multimodal reasoning model, Step3, which will be open-sourced for global enterprises and developers on July 31, 2025 [1] Group 1: Model Development - Step3 features an innovative model architecture and algorithm engineering collaboration, utilizing a MoE architecture with a total parameter count of 321 billion and an active parameter count of 38 billion, achieving significant cost optimization while maintaining high decoding efficiency [2] - Step3 possesses visual perception and complex reasoning capabilities, enabling it to accurately perform cross-domain complex knowledge understanding and analyze the intersection of mathematical and visual information [2] Group 2: Performance and Efficiency - Step3 demonstrated over 70% throughput improvement compared to the industry-leading model DeepSeek-R1 when tested on NVIDIA H800 chips, and can achieve up to 300% efficiency on domestic chips [2] - A core innovation alliance has been established with nearly 10 chip and infrastructure manufacturers to enhance model adaptability and computing efficiency [2][3] Group 3: Strategic Partnerships and Revenue Goals - The alliance includes major domestic chip manufacturers such as Huawei Ascend and others, which have successfully implemented and run Step3 [3] - Jiyue Xingchen has formed a strategic partnership with Shanghai State-owned Capital Investment Co., Ltd., aiming for a revenue target of 1 billion RMB for the year, based on confirmed contract revenues and a strong gross profit margin achieved in the first half of 2025 [3][4] - The company has collaborated with over half of domestic smartphone manufacturers to develop intelligent agents and is exploring applications in various verticals, including finance and retail [4]

Core Viewpoint - Ant Group has successfully utilized domestic chips, including those from Alibaba and Huawei, in conjunction with the mixed expert (MoE) machine learning method to train AI models, achieving a cost reduction of approximately 20% [1] Group 1 - The performance of the new technology is comparable to NVIDIA's H800 chip [1] - Ant Group continues to use NVIDIA chips for AI development but has shifted its latest models to primarily rely on alternatives from AMD and domestic Chinese chips [1] - Ant Group is continuously optimizing for different chips to reduce AI application costs and has made significant progress, with plans to gradually share its findings through open-source initiatives [1] Group 2 - The move is significant in the context of U.S. export restrictions on high-end chips to China, indicating that China has largely overcome U.S. semiconductor sanctions [1] - Ant Group's open-source Ling series model framework and training strategies could promote the accessibility of domestic AI technology, lowering the entry barriers for small and medium-sized enterprises and research institutions [1]