AMD瞄准韩国物理AI市场
AMD正加大力度进军韩国物理AI市场，推出自适应和嵌入式计算平台，计划将业务从半导体设备扩展到汽车领域。
原文深度解读
AMD正加大力度进军韩国物理AI市场，推出自适应和嵌入式计算平台，计划将业务从半导体设备扩展到汽车和机器人领域。文章基于2024年9月2日AMD自适应与嵌入式计算集团高级副总裁兼总经理Salil Raje在首尔汝矣岛办公室接受THE ELEC采访的内容，涉及AMD与韩国半导体设备公司、三星、LG、SK海力士、StradVision、Nota AI的合作，以及其异构计算、开放生态系统和低延迟物理AI技术。
- AMD自适应与嵌入式计算集团高级副总裁兼总经理Salil Raje于9月2日在首尔汝矣岛AMD AECG办公室接受采访，称韩国是战略物理AI市场，因其在工业自动化和汽车技术方面的优势。
- AMD认为物理AI已到来，而非遥远未来技术，AI正深度集成于嵌入式系统，传统机器控制算法正被AI算法取代。
- 三星电子和SK海力士等韩国半导体制造商正使用AI算法自动化制造流程，设备利用AI自动执行关键设置和操作，缩短工艺时间并支持无人操作，减少因操作员技能差异导致的工艺良率变化。
- AMD通过FPGA和自适应SoC等异构计算解决方案为AI自动化做贡献，自适应SoC是异构半导体，内部硬件电路可在芯片制造并安装后重新配置，结合CPU、FPGA和AI引擎。
- AMD与韩国半导体设备公司在工艺设备自动化上合作，涉及前端半导体加工、后端封装和内存测试设备，但未透露具体公司。
- AMD与StradVision和Nota AI合作，StradVision正在为AMD硬件优化其高性能多视觉感知软件，Nota AI正在联合研究AI模型压缩和优化。
供应链影响
- AMD与韩国半导体设备公司的合作可能深化，影响其FPGA和自适应SoC在半导体制造设备中的采用，具体取决于设备自动化需求。
- AMD与三星、LG、SK海力士等大公司的合作可能扩大，可能影响其解决方案在消费电子、存储和汽车等领域的供应链布局，取决于合作进展。
- AMD与StradVision和Nota AI的合作可能推动自动驾驶和AI软件在AMD硬件上的优化，可能影响相关汽车和机器人供应链中的软硬件集成。
- AMD的开放生态系统和机器人合作伙伴网络可能吸引更多合作伙伴，可能改变机器人制造商的模块采购模式，取决于生态系统的扩展。
- AMD的异构单芯片架构（如Ryzen AI Embedded和Versal AI Edge）可能影响低延迟物理AI应用的硬件选择，可能对相关供应链中的计算平台需求产生影响。
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自适应与嵌入式计算集团高级副总裁兼总经理Salil Raje接受THE ELEC采访。（照片：AMD）
AMD正加大力度进军韩国物理人工智能（AI）市场，推出自适应和嵌入式计算平台。该公司旨在通过异构计算和开放生态系统，将业务从半导体设备扩展到车辆和机器人。
AMD自适应与嵌入式计算集团（AECG）高级副总裁兼总经理Salil Raje于9月2日在汝矣岛AMD AECG首尔办公室接受采访时表示，韩国是该公司战略性的物理AI市场，因为韩国在工业自动化和汽车技术方面具有优势。
“物理AI可以在很多方面帮助韩国制造业，因为韩国面临人口萎缩和生产率下降的问题，”Raje说。“与韩国政府和企业的合作可以双赢。”
物理AI是一种下一代AI技术，它感知物理世界，做出判断和决策，并采取直接行动。它可用于医疗保健、医疗应用、无人机和汽车。例如AI机械臂、无人驾驶飞行器、自动驾驶（AD）车辆和自主工厂机械。该技术超越了理解信息，在物理环境中感知和行动。
AMD认为物理AI不是遥远未来的技术，而是已经到来。“现实世界中的许多产品已经在以各种方式使用AI，”Raje说。“AI正深度集成到嵌入式系统中，甚至传统的机器控制算法也被AI算法取代。”嵌入式系统是内置于设备或设备中的专用硬件和软件，用于控制和执行特定功能，这是AMD AECG负责的产品领域。
◇ 半导体设备的AI自动化
Raje讨论AMD与韩国半导体设备公司的合作。（照片：AMD）
包括三星电子和SK海力士在内的韩国半导体制造商正在使用AI算法自动化制造流程。设备使用AI自动执行关键设置和操作，缩短工艺时间并支持无人操作。它还可以减少因操作员技能水平差异导致的工艺良率变化。
AMD通过异构计算解决方案为AI自动化做贡献，包括现场可编程门阵列（FPGA）和自适应片上系统（SoC）。
自适应SoC是一种异构半导体，其内部硬件电路即使在芯片制造并安装到产品中后也可以重新配置。它结合了中央处理器（CPU）、FPGA和AI引擎等组件。类似于人类神经系统，这些组件在它们之间分配工作负载，以最大化性能和功率效率。
“我们继续与韩国半导体设备公司合作，使用以AI为重点的FPGA和自适应SoC产品线进行工艺设备自动化，”Raje说。“我们深度参与前端半导体加工、后端封装和内存测试设备。”他没有透露涉及的设备公司。
传统的机器控制算法是基于规则的。机器根据预定义的统一标准做出决策和操作，导致错误率较高。例如，基于规则的自动检测设备可能无法识别不规则的、非标准的缺陷，可能导致过度杀伤，即良品被错误分类为缺陷品。
◇ 对与三星、LG和SK海力士的合作持开放态度
StradVision的多视觉技术。（照片：StradVision）
AMD正在加强与韩国公司的合作，从工业自动化扩展到汽车和机器人，包括人形机器人。它已经与StradVision和Nota AI合作。自动驾驶感知软件开发商StradVision正在为AMD硬件优化其高性能多视觉感知软件。Nota AI正在联合研究AMD硬件的AI模型压缩和优化。
“与三星、LG和SK海力士等韩国大公司的合作极其重要，它们可以在广泛的应用中部署AMD的各种解决方案，”Raje说。“韩国公司最近的机器人投资对AMD来说是一个巨大的机会，我们对与任何公司合作持开放态度。”
车辆和机器人都需要实时感知周围环境、超低延迟控制和高水平的安全性。即使系统决策稍有延迟，也可能导致严重事故。
◇ 异构单芯片架构实现低延迟物理AI
AMD Ryzen AI嵌入式100系列。（照片：AMD）
AMD的异构半导体技术，包括自适应SoC，在单个芯片内处理物理AI的感知、决策和行动阶段。这最小化了数据移动，有助于防止瓶颈和延迟。支持这种方法的产品和平台包括Ryzen AI嵌入式100系列、Versal AI Edge系列和Vitis统一软件平台。
Ryzen AI嵌入式100系列包括X100和P100。这些产品将基于x86的CPU、高性能集成GPU和神经处理单元（NPU）集成到单个SoC中。CPU管理和协调物理AI的整体操作。GPU分析摄像头和其他传感器收集的大量数据，并为AI工作负载处理视觉信息。NPU以低功耗处理重复的AI推理。集成这三个功能减少了数据通信时间并提高了功率效率。
Versal AI Edge系列在单个SoC中结合了基于Arm的CPU、FPGA和AI引擎。Arm CPU处理顺序和不规则的计算工作负载。FPGA可以立即实现针对系统使用的传感器定制的硬件接口，同时处理电机关节同步和高速信号处理。AI引擎为视觉感知、推理和无线通信信号处理等任务提供轻量级、高速加速。与竞争产品相比，该平台提供高达三倍的计算密度和高达三倍的AI计算效率（以每瓦特每秒万亿次操作（TOPS）衡量）。
Vitis统一平台是一个硬件和软件开发环境，用于实时开发和验证AMD异构架构硬件。传统上，控制FPGA或自适应SoC需要硬件级设计方法，如寄存器传输级（RTL）和硬件描述语言（HDL）。相比之下，Vitis允许使用更高级的C和C++语言一起编程异构核心。Raje监督了该平台的开发和发布。
◇ 开放生态系统提供更大灵活性
Raje强调AMD的开放生态系统是关键优势。（照片：AMD）
AMD还运营机器人合作伙伴网络，这是一个全球协作生态系统，旨在帮助合作伙伴更快、更容易地组装和开发机器人，类似于用乐高积木搭建。AMD提供标准化的高性能技术平台，并验证外部合作伙伴开发的板卡和模块。机器人制造商可以组合兼容的、经过验证的组件和模块，以加速产品商业化。
与竞争对手封闭、垂直整合的锁定方法不同，AMD提供更大的开发灵活性。合作伙伴可以根据需求重新设计系统级模块（SOM）。SOM是一种嵌入式计算模块，将处理器、内存和电源管理等核心计算功能集成到紧凑的电路板上。AMD还为其生态系统中的合作伙伴提供技术和营销支持。
Salil Raje, senior vice president and general manager of AMD’s Adaptive and Embedded Computing Group, speaks in an interview with THE ELEC. (Photo: AMD)
AMD is stepping up its push into Korea’s physical artificial intelligence (AI) market with adaptive and embedded computing platforms. The company aims to expand from semiconductor equipment to vehicles and robots by leveraging heterogeneous computing and an open ecosystem.
Salil Raje, senior vice president and general manager of AMD’s Adaptive and Embedded Computing Group (AECG), said in an interview at AMD AECG’s Seoul office in Yeouido on Sept. 2 that Korea is a strategic physical AI market for the company because of its strengths in industrial automation and automotive technologies.
“Physical AI can help Korean manufacturing in many ways as the country faces a shrinking population and declining productivity,” Raje said. “Cooperation with the Korean government and companies can be a win-win.”
Physical AI is a next-generation AI technology that senses the physical world, makes judgments and decisions, and takes direct action. It can be used in healthcare, medical applications, drones and automobiles. Examples include AI robotic arms, unmanned aerial vehicles, autonomous driving (AD) vehicles and autonomous factory machinery. The technology goes beyond understanding information to perceiving and acting within physical environments.
AMD views physical AI not as a distant-future technology but one that has already arrived. “Many products in the real world are already using AI in various ways,” Raje said. “AI is becoming deeply integrated across embedded systems, with even traditional machine-control algorithms being replaced by AI algorithms.” Embedded systems are dedicated hardware and software built into devices or equipment to control and perform specific functions, a product area overseen by AMD’s AECG.
◇ AI Automation for Semiconductor Equipment
Raje discusses AMD’s cooperation with Korean semiconductor equipment companies. (Photo: AMD)
Korean semiconductor manufacturers including Samsung Electronics and SK hynix are using AI algorithms to automate manufacturing processes. Equipment uses AI to automatically perform key settings and operations, shortening process times and supporting unmanned operations. It can also reduce variations in process yields caused by differences in operator skill levels.
AMD is contributing to AI automation with heterogeneous computing solutions including field-programmable gate arrays (FPGAs) and adaptive systems-on-chip (SoCs).
An adaptive SoC is a heterogeneous semiconductor whose internal hardware circuitry can be reconfigured even after the chip has been manufactured and installed in a product. It combines components including a central processing unit (CPU), FPGA and AI engines. Similar to the human nervous system, the components divide workloads among themselves to maximize performance and power efficiency.
“We are continuing to work with Korean semiconductor equipment companies on process equipment automation using AI-focused FPGA and adaptive SoC product lines,” Raje said. “We are deeply involved in front-end semiconductor processing, back-end packaging and memory test equipment.” He did not identify the equipment companies involved.
Traditional machine-control algorithms were rule-based. Machines made decisions and operated according to predefined, uniform criteria, resulting in higher error rates. For example, rule-based automated inspection equipment can fail to identify irregular, non-standard defects, potentially leading to overkill, in which good products are incorrectly classified as defective.
◇ Open to Partnerships With Samsung, LG and SK hynix
StradVision’s multi-vision technology. (Photo: StradVision)
AMD is strengthening cooperation with Korean companies beyond industrial automation into automotive and robotics, including humanoid robots. It is already working with StradVision and Nota AI. Autonomous-driving perception software developer StradVision is optimizing its high-performance multi-vision perception software for AMD hardware. Nota AI is jointly researching AI model compression and optimization for AMD hardware.
“Cooperation with major Korean companies such as Samsung, LG and SK hynix, which can deploy AMD’s various solutions across a broad range of applications, is extremely important,” Raje said. “Recent robotics investments by Korean companies represent a tremendous opportunity for AMD, and we remain open to working with any company.”
Both vehicles and robots require real-time perception of their surroundings, ultra-low-latency control and high levels of safety. Even a brief delay in system decision-making can lead to serious accidents.
◇ Heterogeneous Single-Chip Architecture Enables Low-Latency Physical AI
AMD Ryzen AI Embedded 100 Series. (Photo: AMD)
AMD’s heterogeneous semiconductor technologies, including adaptive SoCs, handle the sensing, decision-making and action stages of physical AI within a single chip. This minimizes data movement and helps prevent bottlenecks and latency. Products and platforms supporting this approach include the Ryzen AI Embedded 100 Series, Versal AI Edge Series and Vitis unified software platform.
The Ryzen AI Embedded 100 Series consists of the X100 and P100. The products integrate an x86-based CPU, high-performance integrated GPU and neural processing unit (NPU) into a single SoC. The CPU manages and coordinates the overall operation of physical AI. The GPU analyzes large volumes of data collected by cameras and other sensors and processes visual information for AI workloads. The NPU handles repetitive AI inference at low power. Integrating the three functions reduces data communication time and improves power efficiency.
The Versal AI Edge Series combines Arm-based CPUs, FPGAs and AI engines in a single SoC. The Arm CPU handles sequential and irregular computing workloads. The FPGA can immediately implement hardware interfaces tailored to the sensors used by a system while handling motor-joint synchronization and high-speed signal processing. The AI engines provide lightweight, high-speed acceleration for tasks including visual perception, inference and wireless communications signal processing. The platform delivers up to three times the compute density and up to three times the AI compute efficiency, measured in tera operations per second (TOPS) per watt, compared with competing products.
The Vitis unified platform is a hardware and software development environment for developing and validating AMD heterogeneous-architecture hardware in real time. Controlling FPGAs or adaptive SoCs traditionally required hardware-level design methods such as register-transfer level (RTL) and hardware description languages (HDLs). Vitis, by contrast, allows heterogeneous cores to be programmed together using higher-level C and C++ languages. Raje oversaw the platform’s development and launch.
◇ Open Ecosystem Offers Greater Flexibility
Raje highlights AMD’s open ecosystem as a key strength. (Photo: AMD)
AMD also operates a Robotics Partner Network, a global collaborative ecosystem designed to help partners assemble and develop robots more quickly and easily, similar to building with Lego blocks. AMD provides standardized, high-performance technology platforms and validates boards and modules developed by external partners. Robot manufacturers can combine compatible, validated components and modules to accelerate product commercialization.
Unlike competitors’ closed, vertically integrated lock-in approaches, AMD provides greater development flexibility. Partners can redesign systems-on-module (SOMs) to suit their requirements. A SOM is an embedded computing module that integrates core computing functions including processors, memory and power management onto a compact circuit board. AMD also provides technical and marketing support to partners participating in its ecosystem.