推理芯片市场，HBM迎来了挑战者

Core Viewpoint - The article highlights the rapid growth of the AI market, emphasizing SK Hynix's emergence as a leading DRAM supplier, surpassing Samsung, driven by its HBM technology and the evolving demands of AI applications [1][3]. Group 1: SK Hynix and HBM Technology - SK Hynix achieved revenue of $16.23 billion and a profit of $5.1 billion in Q2 2025, marking a 69.8% year-over-year growth, making it the world's top DRAM supplier [1]. - HBM, which accounts for 77% of SK Hynix's revenue, utilizes vertical stacking and TSV technology to enhance memory bandwidth, playing a crucial role in training large AI models [1][3]. - The cost pressures associated with HBM have led to a shift towards GDDR memory for AI inference applications [1][3]. Group 2: Evolution of AI - The article describes the transition from AI 1.0, characterized by simpler applications like voice assistants and recommendation engines, to AI 2.0, which includes large language models (LLMs) capable of understanding complex inputs and generating diverse outputs [2][3]. - The scale of AI models has significantly increased, with parameters in models like Chat GPT-4 reaching 1.76 trillion, highlighting the growing demand for memory bandwidth and capacity [3]. Group 3: GDDR Memory Advantages - GDDR memory, originally designed for GPUs, offers high data transfer rates and is now favored for AI inference due to its cost-effectiveness and performance [4][5]. - GDDR7, with a data rate of 192 GB/s and a chip density of 32 Gb, is positioned as a suitable option for edge networks and IoT devices, providing high bandwidth at lower costs compared to HBM [4][5][8]. - GDDR7's performance is particularly notable, offering 128 GB/s bandwidth, which is more than double that of alternative solutions, making it ideal for AI inference applications [8]. Group 4: Rambus and GDDR7 Controller - Rambus has developed the industry's first GDDR7 memory controller IP, which supports data rates up to 40 Gbps and provides 160 GB/s of usable bandwidth per device [12]. - The GDDR7 controller is designed for high memory throughput and low latency, utilizing advanced scheduling algorithms to optimize bus efficiency [13]. - Rambus's expertise in signal integrity and power integrity positions it as a key player in the AI chip market, enhancing the performance of GDDR7 memory in advanced applications [11][12].