AMD董事长苏姿丰：“将在韩国开设AI中心并招聘研究人员”
AMD董事长苏姿丰于10月7日宣布，公司计划年内在韩国设立“AMD人工智能（AI）研究中心”，并招聘数百名研究人员。
原文深度解读
文章报道AMD董事长苏姿丰于10月7日在首尔宣布，计划年内在韩国设立“AMD人工智能（AI）研究中心”（COE），并招聘数百名AI研究人员；这是韩国科学技术信息通信部与AMD在7月总统李在明旧金山AI峰会期间交换谅解备忘录（MOU）的后续措施。文章涉及韩国政府、AMD、14家韩国AI半导体与软件企业，以及三星电子、SK海力士等合作方；关键数字包括10月7日、原定10月8日提前一天、14家企业、数百名研究人员、年内设立、下月指导会议等。与电子元器件供应链的关系在于：AMD拟通过ROCm开放软件平台支持GPU及其他加速器（包括FuriosaAI、HyperAccel、Mobilint的NPU），并讨论将AMD CPU、GPU与韩国NPU组合为异构AI计算系统，可能影响AI芯片、加速器、存储与相关元器件的需求结构与生态合作。
- AMD董事长苏姿丰10月7日在首尔宣布，计划年内在韩国设立“AMD AI研究中心”并招聘数百名AI研究人员；这是韩国科学技术信息通信部与AMD在7月总统李在明旧金山AI峰会期间交换MOU的后续措施。
- 韩国AI COE正进行工作层磋商，目标年内设立；政府决定选址，AMD负责硬件等投资，具体细节计划在下月指导会议准备，组织开放后预计成为AMD在亚太地区首个COE。
- AMD既有COE通常是在当地投资或存在执行大规模项目的战略需求时创建；韩国中心因由MOU触发、没有万亿韩元级设施投资或具体合同订单确认而受到关注。
- 苏姿丰强调与韩国企业合作扩展AMD开放生态，传达通过软件平台ROCm不仅可选择GPU，也可选择其他加速器的战略方向；其他加速器包括当天参会的FuriosaAI、HyperAccel、Mobilint的国产NPU。
- ROCm是AMD的开放计算软件平台，由驱动、编译器、库和开发工具组成，以开源软件提供高可访问性；文章将其与Nvidia的CUDA类比，称CUDA是在其自有GPU上运行AI模型的开发环境，ROCm在AMD芯片上承担该角色。
- 与会者讨论将AMD的CPU和GPU与国产NPU捆绑为单一系统的异构AI计算；判断随着市场中心从训练转向推理，通过连接不同半导体和软件降低推理成本与功耗，比单一芯片性能更重要；双方同意将其落地为实际基础设施，并通过国内验证发展为可用于全球市场的参考模型，同时提到AMD与MangoBoost为推理服务构建的云协作案例。
供应链影响
- 若韩国COE按计划年内设立并招聘数百名研究人员，可能增加对AMD GPU、CPU及相关服务器/加速器硬件的本地需求，但具体规模取决于政府选址、AMD投资额和下月指导会议结果。
- ROCm若支持FuriosaAI、HyperAccel、Mobilint等国产NPU，可能推动韩国AI加速器与AMD平台之间的软硬件适配和生态合作，进而影响相关芯片、板卡与系统集成的供应链协作方式。
- 异构AI计算将AMD CPU、GPU与国产NPU组合为单一系统，可能带动对互连、内存、封装、电源与散热等配套元器件的差异化需求，但实际影响取决于参考模型验证和商业化落地进度。
- 市场中心从训练转向推理的判断，可能使供应链更关注推理成本与功耗优化，而非单一芯片峰值性能；这可能影响加速器、存储和系统级方案的选型方向，但取决于客户采用与全球市场验证情况。
- 苏姿丰计划与三星电子、SK海力士管理层会面，可能涉及AI芯片与存储供应合作；潜在供应链影响取决于会谈内容及后续是否形成具体合同或订单，原文未披露相关细节。
系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。
AMD董事长苏姿丰于10月7日在首尔钟路区四季酒店与韩国科学技术信息通信部第二次官柳济明讨论韩国AI的未来。照片由Lee Dong-geun拍摄，foto@etnews.com
AMD董事长苏姿丰于10月7日宣布，公司计划年内在韩国设立“AMD人工智能（AI）研究中心”，并招聘数百名研究人员。这是韩国科学技术信息通信部与AMD在7月总统李在明旧金山AI峰会期间交换谅解备忘录（MOU）的后续措施。
韩国科学技术信息通信部第二次官柳济明与苏姿丰当天在首尔四季酒店会面，并举行了有14家韩国AI半导体和软件（SW）企业参加的圆桌会议。由于苏姿丰的日程，原定10月8日的活动提前了一天，因此由柳济明次官代替副总理裴庆勋出席。
在与柳济明的会面中，苏姿丰表示：“我们正在推动设立韩国AI‘卓越中心（COE）’，并计划招聘数百名AI研究人员。”
COE是企业等主体聚集特定领域研究、验证和人才培养功能的核心组织。AMD的韩国COE正在进行工作层磋商，目标是在年内设立。政府将决定选址，AMD将负责硬件等投资。具体细节计划在下月的指导会议上准备，组织开放后，预计将成为AMD在亚太地区的首个COE。
AMD既有的COE是在AMD此前进行当地投资，或存在执行大规模项目的战略需求时创建的。韩国中心因由谅解备忘录触发、没有万亿韩元级设施投资或具体合同订单确认而受到关注。
韩国科学技术信息通信部第二次官柳济明与AMD董事长苏姿丰于10月7日在首尔钟路区四季酒店举行的AI半导体企业圆桌会议上与参会者合影。照片由Lee Dong-geun拍摄，foto@etnews.com
苏姿丰还强调与韩国企业合作，以扩展AMD的开放生态。她传达了战略方向：通过其软件平台“ROCm（Radeon Open Compute）”，不仅可以选择图形处理器（GPU），还可以选择其他加速器。其他加速器包括当天参会的FuriosaAI、HyperAccel和Mobilint生产的国产神经处理单元（NPU）。
ROCm是AMD的开放计算软件平台。如果Nvidia的“CUDA”是在其自有图形处理器上运行人工智能模型的开发环境，那么ROCm在AMD芯片上承担这一角色。它由驱动程序、编译器、库和开发工具组成，并作为开源软件提供高可访问性。
当天，参会者还讨论了将AMD的中央处理器和图形处理器与国产神经处理单元捆绑为单一系统的异构AI计算。随着市场中心从训练转向推理，人们判断，通过连接不同半导体和软件来降低推理成本和功耗，已经比单一芯片的性能更重要。
双方同意将其落实到实际基础设施中，并通过国内验证，发展为可在全球市场利用的参考模型。为增加韩国AI计算能力，以便国内初创企业和用户使用，相关投资也在进行中。AMD与MangoBoost为推理服务目的构建的云协作案例也被提及。
苏姿丰解释说：“让用户轻松选择他们想要的任何硬件，无论是图形处理器还是加速器，是AMD的理念，ROCm就是这条路径。”“我们也对连接中央处理器和各种加速器的合作持开放态度。”
同一天，苏姿丰前往龙山区Dragon City，与韩国游戏玩家进行交流。之后，她计划与合作伙伴企业管理层举行会议，包括三星电子副会长全永铉和SK海力士社长郭鲁正。
苏姿丰强调：“我来（韩国）是为了见合作伙伴，我将见三星，也会见SK海力士。”“AI是我职业生涯中发展最快的领域，访问韩国并定期与合作伙伴沟通，对于应对市场很重要。”
Lisa Su, chair of AMD, talks about the future of AI in South Korea with Ryu Je-myeong, second vice minister of science and ICT, at the Four Seasons Hotel in Jongno-gu, Seoul, on Oct. 7. Photo by Lee Dong-geun foto@etnews.com
Lisa Su, chair of AMD, announced on Oct. 7 that the company plans to establish an “AMD Artificial Intelligence (AI) Research Center” in South Korea within the year and hire hundreds of research personnel. This is a follow-up measure to the memorandum of understanding (MOU) exchanged between the Ministry of Science and ICT and AMD on the occasion of President Lee Jae-myung's San Francisco AI Summit in July.
Ryu Je-myeong, second vice minister of science and ICT, and Su met at the Four Seasons Hotel in Seoul on the day and held a roundtable in which 14 domestic AI semiconductor and software (SW) companies participated. As the event, originally scheduled for Oct. 8, was moved forward by a day due to Su's schedule, Vice Minister Ryu attended instead of Deputy Prime Minister Bae Kyung-hoon.
In her meeting with Ryu, Su said, “We are pushing to establish a Korean AI 'Center of Excellence (COE)' and plan to hire hundreds of AI research personnel.”
A COE is a core organization where companies and other entities aggregate research, validation, and talent cultivation functions for specific fields. AMD's Korean COE is undergoing working-level consultations with the goal of being established within the year. The government will determine the location, while AMD will be in charge of investments such as hardware. Specific details are scheduled to be prepared at next month's steering meeting, and once the organization opens, it is expected to become AMD's first COE in the Asia-Pacific region.
AMD's existing COEs were created when AMD had previously invested locally or when there was a strategic need to perform large-scale projects. The Korean center is drawing attention as it is the first case being promoted triggered by a memorandum of understanding without trillion-won-scale facility investments or confirmation of specific contract orders.
Ryu Je-myeong, second vice minister of science and ICT, and Lisa Su, chair of AMD, pose for a photo with attendees during an AI semiconductor company roundtable at the Four Seasons Hotel in Jongno-gu, Seoul, on Oct. 7. Photo by Lee Dong-geun foto@etnews.com
Su also emphasized collaboration with Korean companies to expand AMD's open ecosystem. She conveyed a strategic direction to allow the selection of not only graphics processing units (GPUs) but also other accelerators through its software platform 'ROCm (Radeon Open Compute)'. Other accelerators encompass domestic neural processing units (NPUs) produced by FuriosaAI, HyperAccel, and Mobilint, which participated that day.
ROCm is AMD's open computing software platform. If Nvidia's 'CUDA' is a development environment for running artificial intelligence models on its own graphics processing units, ROCm serves that role on AMD chips. It consists of drivers, compilers, libraries, and development tools, and offers high accessibility as open-source software.
On this day, attendees also discussed heterogeneous AI computing that bundles AMD's central processing units and graphics processing units with domestic neural processing units into a single system. As the market center shifts from training to inference, it has been judged that lowering inference costs and power consumption by connecting different semiconductors and software has become more important than the performance of a single chip.
Both sides agreed to implement this into actual infrastructure and develop it into a reference model that can be utilized in the global market through domestic verification. Investments are also underway to increase AI computing capacity in South Korea for domestic startups and users to utilize. The cloud collaboration case between AMD and MangoBoost, constructed for the purpose of inference serving, was also mentioned.
“Making it easy for users to choose whichever hardware they want, whether graphics processing units or accelerators, is AMD's philosophy, and ROCm is that pathway,” Su explained. “We are also open to collaborations that connect central processing units and various accelerators.”
On the same day, Su moved to Dragon City in Yongsan-gu to have time to communicate with Korean gamers. Afterward, she is scheduled to carry out meetings with management of partner companies, including Jun Young-hyun, vice chairman of Samsung Electronics Co., and Kwak Noh-jung, president of SK Hynix Inc.
“I came (to South Korea) to meet partners, and I will meet Samsung and also SK Hynix,” Su emphasized. “AI is the fastest-moving field in my career, and visiting South Korea and regularly communicating with partners is important in responding to the market.”