科技动态：SRAM- 一种全新的 AI 推理范式-Tech Bytes-SRAM

Summary of Key Points from the Conference Call Industry Overview - The memory market for AI inference is transitioning towards a hybrid approach, with SRAM gaining traction for latency-sensitive workloads, while HBM remains dominant for capacity-heavy tasks [1][3][4]. Core Insights - SRAM and LPU Architecture: NVIDIA is expected to introduce an inference chip based on LPU architecture that utilizes on-chip SRAM, enhancing speed for AI inference applications. This architecture is designed for sequential speed rather than massive parallel processing [2][12]. - Performance Comparison: SRAM provides ultra-low latency and immediate data availability, making it suitable for speed-critical applications, while DRAM offers higher capacity and lower cost per bit, serving as the backbone for high-capacity external memory [11][21]. - Cost Efficiency: The cost per token generated is a critical metric for inference, with LPUs capable of generating tokens efficiently at full compute capacity, thus lowering energy costs compared to GPUs [10][12]. Implications for the Market - Partnership of SRAM and HBM: The relationship between SRAM and HBM is characterized as a division of labor, where SRAM is used for speed-sensitive applications and HBM for scalable memory capacity. This partnership is crucial for the evolving landscape of AI applications [3][4]. - Supply Chain Advantages: The LPU architecture can bypass supply chain bottlenecks associated with HBM, allowing for effective designs even on older foundry process nodes [4]. Investment Insights - Stock Recommendations: Samsung Electronics is highlighted as a top pick due to its HBM4 qualification, SRAM capabilities, and foundry optionality. SK hynix is also recommended as an overweight investment [5]. - Market Corrections: Recent stock price corrections (20% WTD vs. KOSPI -17%) present buying opportunities, as historical trends indicate that stock prices often recover beyond fundamental growth trajectories [5]. Risks and Considerations - Potential Risks: Investors should be aware of risks such as demand fluctuations, rising competition, and elevated inventories among cloud and smartphone customers, which could impact stock performance [32][30]. - Technological Developments: Emerging technologies like MRAM, ReRAM, and eDRAM may recalibrate the memory landscape, but SRAM and DRAM are expected to remain foundational in high-performance computing systems [25]. Additional Insights - LPU Workload Suitability: LPUs are ideal for low-latency applications such as financial transactions and conversational AI, while they are less suited for ultra-large models and batch processing tasks that benefit from GPU clusters [27][12]. - Future Outlook: The memory market is expected to evolve with advancements in technology, and the integration of SRAM and HBM will play a significant role in shaping future AI systems [25][4].