"The advancement of artificial intelligence (AI) is no longer just a graphics processing unit (GPU) issue. With AI agents (assistants) emerging, the importance of semiconductors related to central processing units (CPUs), networks, and software overall is rising. The rise of physical AI is affecting the entire semiconductor ecosystem. AMD aims to provide end-to-end solutions, including environments not consolidation to the cloud."
Lee Jae-hyeong, head of commercial sales at AMD Korea, said this at an "AI solutions briefing" for Korean media held on the 1st at the AMD AECG Seoul office in Yeongdeungpo-gu, Seoul. He said, "The center of the Generative AI market is spreading from conventional training to inference, and then to agentic and physical AI," and added, "AMD expects this shift to lead to a 'full-stack' competition that bundles CPUs, GPUs, networks, and edge semiconductors together." At the briefing, Lee and Kim Hyuk, Asia-Pacific (APAC) tech lead (executive director) at AMD Adaptive and Embedded Computing Group (AECG), took the stage as presenters.
Lee and Kim said the AI Semiconductor race can no longer be explained solely as a performance contest of a single GPU. They noted that the structure of the AI infrastructure market is changing rapidly.
These changes are also evident in AMD's results. AMD's first-quarter revenue this year was $10.253 billion (about 16 trillion won), up 38% from a year earlier. The data center institutional sector led growth with $5.8 billion, up 57% from a year earlier. The client and gaming institutional sector also rose 23% to $3.6 billion over the period, and the embedded institutional sector climbed 6% to $873 million. In particular, client revenue jumped 26% on demand for Ryzen processors, and gaming revenue rose 11% on demand for Radeon GPUs. Revenue grew not only in data centers but also in PCs, graphics, and industrial and edge semiconductors.
◇ "Rival products must be allowed in"... AMD targets market with open AI
Lee cited an "open ecosystem" as the core of AMD's AI strategy. While Nvidia has built a closed full-stack strategy around its GPUs and CUDA software ecosystem, AMD aims for a structure where customers can combine CPUs, GPUs, and networking equipment as needed.
Lee said, "Even in the CPU field where AMD is strong, if a customer wants, we support building AI systems using rival products," adding, "For GPUs as well, if customers want Nvidia products included, AMD's goal is to enable AI rack configurations—this is the open ecosystem we pursue." He added, "An open system that allows rivals in will help AI progress more in the long run, and we believe it will actually draw more customers."
Along with an open ecosystem, AMD is making ▲leadership compute engines and ▲full-stack solutions the main pillars of its AI strategy. With a broad AI portfolio—including EPYC CPUs, Instinct GPUs, Pensando networking products, Versal adaptive system-on-chips (SoCs), and Ryzen and Radeon—it says it can meet diverse customer needs.
Lee explained, "The performance of each individual GPU matters, but when you link 1,000 of them, you don't get the performance of 1,000 as is," adding, "What's needed to extract performance as close to 1,000x as possible is a network solution."
Large-scale AI training and inference tie multiple GPUs together. At that point, data movement speed and latency between GPUs determine overall performance. Lee said that using the company's AI solutions—such as the Instinct MI350 series, ROCm 7, and the next-generation AI rack "Helios"—allows you to design systems that maximize GPU performance.
◇ "Agents use computing resources like employees"... CPU role grows again
Lee said agentic AI is driving up CPU demand. Agentic AI is not a structure where a single large language model (LLM) only generates answers. It loads multiple systems—such as databases, web services, caches, middleware, and application programming interfaces (APIs)—and splits up the work. In this process, GPUs handle large-scale parallel computation, but CPUs are heavily relied upon to allocate tasks, prepare data, and operate the system.
Lee said, "The data pre-processing and post-processing required to run agent AI are based on CPU resources," adding, "Workplaces are forming where one employee has 5 to 20 AI agents, like employees, and leads a team, so CPU demand will explode further going forward."
AMD CPUs are cited as well-suited parts for this market shift. According to market research firm Mercury Research, AMD's x86 server CPU revenue share in the first quarter of this year was 46.2%, a record high. The share by shipments was 33.2%. This means that while AMD lags Intel in total server CPU shipments, it has rapidly raised its revenue share by focusing on high-priced, high-core products.
◇ "You can't delay a left-turn decision by 2 seconds"... Edge AI spreads to robots and cars
Executive director Kim Hyuk focused his presentation on AMD's Edge AI solutions. He said, "If cloud AI is a structure that trains and infers large models in data centers, Edge AI goes into devices that must make decisions on the spot, like cars, robots, and factory equipment," citing "real-time response" and "power efficiency" as the core conditions for Edge AI.
Kim said, "Edge AI solutions are semiconductors that operate in devices like self-driving cars," adding, "If it takes two seconds to answer whether to make a left or a right turn, the moment has already passed." He continued, "You have to respond to road conditions in real time—if the response is slow, the product has no value," adding, "That's why response time matters as much as performance in Edge AI."
Kim views physical AI—especially robots—as structurally similar to cars. Robots take in data from cameras, lidar, and tactile sensors to determine the position and status of objects, then move motors and actuators. Kim said, "You receive camera footage to recognize it, actuate the actuators to decide how far to go and grab, and then determine with tactile sensors whether it has been grabbed," adding, "It's a structure similar to self-driving cars, which recognize road conditions with multiple sensors and make judgments."
He explained that to build such systems, balance among CPUs, GPUs, and neural processing units (NPUs) is important. GPUs suit compute-heavy tasks like large AI models or vision-language models (VLMs), and NPUs are advantageous for processing continuous vision data from cameras at low power. CPUs handle pre-processing to convert raw sensor inputs into forms AI models can process, and post-processing to link AI outputs to actual operations.
Kim said, "AMD has strengths in CPUs and FPGAs, which are essential for pre-processing and post-processing, and that's drawing attention in the market." FPGAs are semiconductors whose circuit structures can be reconfigured to accelerate specific tasks, and they are often used in industrial equipment where camera resolutions or sensor configurations change.