摩尔线程创始人张建中接受中新社采访,分享其辞去英伟达高管职务创立国产GPU企业的初衷,强调全功能GPU是解决‘卡脖子’问题的根本途径,透露公司已量产五颗芯片、推出五代架构,并计划启动10万卡规模集群攻关以支撑万亿参数级大模型训练。
AI-generated summary
张建中曾任英伟达全球副总裁、中国区总经理,2020年辞职创立摩尔线程,旨在填补国内高端GPU依赖进口的空白。
中新社北京9月13日电 题:对话摩尔线程张建中:打造国产GPU,一家民企的造“芯”路
中新社记者 刘亮
“倘若无人迈出这一步,这个赛道或将长期缺席中国身影。”摩尔线程创始人、董事长兼首席执行官张建中近日接受中新社记者采访,分享其创业初衷和对国产GPU(通用计算单元)发展的思考。
2020年,时任英伟达全球副总裁、中国区总经理的张建中辞去高管职务,创立摩尔线程。彼时,全球人工智能(AI)发展正处于爆发前夜,中国高端GPU绝大部分依赖进口。“创立摩尔线程,主要是为了填补国内这块空白。”张建中说。
为填补这一空白,摩尔线程选择从“全功能GPU”入手。张建中判断,未来的计算场景需要一个综合具备人工智能计算、图形渲染、科学计算及物理仿真、超高清视频编解码能力的加速计算平台。
“从国家科技发展与产业安全的角度来看,这也是解决‘卡脖子’问题的根本途径。”张建中强调,如果不去做全功能GPU,未来在发展过程中依然需要去国外采购图形计算芯片等,产业安全依旧会面临风险。
在现有GPU路线中,“全功能GPU”是业内公认技术挑战难度最大的路线之一。“为了实现这个愿景,整个团队必须耐得住寂寞,耐心完成全功能GPU在各个环节的技术突破,无论是3D图形计算、大模型训练推理,还是科学计算和物理仿真等,都需要全面发展。”张建中说。
过去近6年,摩尔线程团队逐步推进研发与产品落地。正式运营18个月后,摩尔线程成功研发出首颗全功能GPU芯片,到如今已成功量产五颗芯片、推出五代架构、构建起覆盖“云-边-端”的全栈产品布局……摩尔线程持续攻坚算力难题,探索建设自主全栈算力体系。
张建中介绍,目前摩尔线程夸娥智算集群已实现从千卡向万卡级规模的扩展,能够支撑万亿参数级别的模型训练,多项关键指标达到国际主流水平。
摩尔线程今年计划启动10万卡规模集群的攻关。张建中介绍,训练万亿乃至未来的10万亿参数大模型,需要依赖万卡或10万卡级别的算力规模。只有保障这些设备在3至6个月的训练周期内稳定运行,才能避免训练时长过长给大模型企业带来负担。
“我们希望通过接下来的技术攻关,真正解决行业内算力训练平台被‘卡脖子’的痛点。”张建中说,未来公司将建设更大规模的智算集群,以支撑AI基础设施发展。
摩尔线程的发展,是当前中国众多民营科技企业探索自主创新的一个缩影。张建中表示,一家民营企业无法单独加速整个产业的发展,只有大家在不同赛道、不同技术路线上共同努力,国家的科研力量才能持续发展。
就GPU行业而言,张建中表示,当前,国内GPU行业发展迅速,企业数量渐多且技术路线各异,这是一个良性发展态势。但无论是哪家企业,想要取得突破,都要共同解决行业发展面临的共性难题。
他强调,国产芯片要想最终取得市场认可,必然要经历从“性价比不如人”到“持平”,再到“超越”的过程,最终实现更好的性能、更具竞争力的性价比和更优的服务。“这也是中国制造走向世界的必经之路。”
“客观上,目前国产GPU芯片距离世界先进水平仍有差距,还需花费更多时间努力追赶。”张建中说。
展望未来,他认为,通过产业合作伙伴的多方努力,未来国产GPU芯片有望逐步达到与国外产品相当的水平。随着生产能力和供应链管理水平的进一步提升,国产GPU将逐步具备与全球头部企业同台竞技的实力。(完)
AI outlook — possibilities, not facts
摩尔线程将启动10万卡规模智算集群攻关,以支撑万亿乃至10万亿参数大模型训练
Likely · Within months
国产GPU芯片将逐步达到与国外产品相当的水平
Possible · Within years

A TrendForce report shows that in the second quarter of 2026, the combined revenue of the world's top ten fabless IC design companies increased by 73% year-on-year to US$141.45 billion. NVIDIA ranked first with revenue of US$91.16 billion. AMD benefited from agent AI applications and rose to third place. Qualcomm retreated to fourth place due to the weakening of its mobile phone business.
The 2026 Inclusion Bund Conference concluded in Shanghai on September 12. Focusing on the theme of "Creating a New AI Economy", it attracted more than 670 guests and more than 300 companies, with more than 78,000 on-site attendees and 23 million online viewers. The conference launched more than 50 technology products, set up industry docking, venture capital activities and an AI art festival, promoted more than 80 cooperation intentions, and held a technology talent recruitment fair to provide more than 1,500 positions.

The Department of Justice of the Guangxi Zhuang Autonomous Region and the Wuzhou Municipal People's Government jointly held a press conference to introduce the 2026 China (Guangxi)-ASEAN "AI + Law" Application Innovation Competition. The competition focuses on the pain points of cross-border legal services and promotes the integration of artificial intelligence and legal services. It has three major tracks: domestic applications, college students, and ASEAN applications. Registration is open to mainland China, Hong Kong, Macao, Taiwan, and ASEAN countries. It has been launched on August 31.

Apple released its first foldable screen phone iPhone Duo on the 10th, Taiwan time. Lai Yiwei, an assistant professor in the Department of Electrical Engineering at National Taiwan Normal University, pointed out that the internal and external screen ratio is √2:1. The design was inspired by Lichtenberg’s proportional scaling paper specifications. He emphasized that this ratio can achieve seamless filling of the screen in different forms. It is not an exact value but an engineering compromise.

Chen Yiru, director of government relations for Shield AI in Asia, said that Taiwan plays a key role in building a drone supply chain that is not dependent on China. Currently, the proportion of Taiwanese materials in its products has increased from less than 1% two years ago to 16%, and is expected to reach 25%. He also emphasized long-term investment in Taiwan, deepening local cooperation and technology transfer to enhance Taiwan's defense resilience.

The Chinese Academy of Sciences cooperated with the American company Shield AI to complete the autonomous flight verification of three Kingbee III drones a month ago, using Hivemind AI pilot technology to achieve autonomous reconnaissance, take-off and landing missions, reducing the operator's burden and improving defense resilience.