[Edaily Reporter Kong Chi-Yu ] AMD, often referred to as the “number two” player in the artificial intelligence (AI) semiconductor market, has unveiled a roadmap to introduce next-generation AI accelerators with improved performance every year. The amount of High-Bandwidth Memory (HBM) is expected to increase significantly with each new product generation. Consequently, SamsungElectronics, which supplies HBM to AMD, is also expected to benefit from this expansion in supply.

Lisa Su, CEO of AMD, stated at “Advancing AI 2026” held in San Francisco, U.S., on the 23rd (local time), “We will introduce a new generation of ‘Instinct’ AI accelerators every year,” adding, “With each new generation of Instinct, we will provide more memory, more bandwidth, and significantly greater computing performance.”

On that day, AMD officially unveiled its 6th-generation server central processing unit (CPU), “EPYC,” and the server AI accelerator “Instinct MI455X.” The company also officially launched the Helios server rack based on these products. CEO Su announced that day, “Helios has entered full-scale mass production.” Each Helios unit is equipped with 72 MI455X accelerators and 18 EPYC CPUs. It also features 31 terabytes (TB) of 6th-generation HBM4 memory.

As the AI trend has recently shifted toward the inference stage, demand for AI accelerators is also growing exponentially. The transition to agentic AI—where AI agents repeatedly perform inference, invoke tools, and access data through numerous steps—is driving significant growth in the AI accelerator market. AMD had previously projected last year that the AI accelerator market would reach $500 billion (approximately 733 trillion won) by 2028, but has revised this forecast upward just one year later. CEO Soo predicted, “The AI accelerator market will reach $1.4 trillion (approximately 2,054 trillion won) by 2030.”

With the MI455X accelerator and Helios server racks entering full-scale mass production, SamsungElectronics—which supplies HBM for AMD’s AI chips—is also expected to benefit. Last March, SamsungElectronics and AMD signed a memorandum of understanding (MOU) at SamsungElectronics’ Pyeongtaek plant, outlining plans to integrate Samsung’s HBM into AMD’s next-generation AI accelerators, including the Instinct MI455X.

Given that AMD has committed to releasing AI accelerators with upgraded performance every year, demand for SamsungElectronics’ next-generation HBM is also expected to grow steadily. In its roadmap released today, AMD announced plans to launch the “MI500” GPU series next year and the “MI600” series in 2028. The MI500 will feature 7th-generation HBM4E. The MI455X, released this year, was equipped with 432 gigabytes (GB) of HBM4, and the HBM capacity in the MI500 is expected to exceed that.

Meanwhile, NVIDIA, a leader in AI semiconductors, will equip its Vera Rubin chip with 288 GB of HBM4 this year. NVIDIA has previously revealed a roadmap stating that it will incorporate 1TB of HBM4E into the Vera Rubin Ultra, set for release next year. An industry official stated, “As demand for AI servers grows rapidly, the HBM capacity and performance incorporated into next-generation AI chips are also increasing,” adding, “Supply opportunities for memory manufacturers, including SamsungElectronics, will also grow accordingly.”