公众号记得加星标⭐️，第一时间看推送不会错过。 编者按 日前，由Xiaoyu Ma和David Patterson联合署名的文章《Challenges and Research Directions for Large Language Model Inference Hardware》正式发布。这篇文章被发布以后，引起了广 泛关注。文章中，作者围绕LLM推理芯片的挑战以及解决方案，给出了建议。 以下为文章正文： 大型语言模型 (LLM) 推理难度很高。底层 Transformer 模型的自回归解码阶段使得 LLM 推理与训 练有着本质区别。受近期人工智能趋势的影响，主要挑战在于内存和互连，而非计算能力。 为了应对这些挑战，我们重点介绍了四个架构研究方向：高带宽闪存，可提供 10 倍内存容量，带宽 堪比 HBM；近内存处理和 3D 内存逻辑堆叠，可实现高内存带宽；以及低延迟互连，可加速通信。 虽然我们的研究重点是数据中心人工智能，但我们也探讨了这些方案在移动设备上的应用。 引言 当一位作者于 1976 年开始其职业生涯时，计算机体系结构会议上约 40% 的论文来自业界。到 2025 年 ISCA 会议时，这一 ...

Group 1 - The core viewpoint of the articles emphasizes the resurgence of AI concept stocks in the A-share market, driven by a focus on computing power and applications, particularly in storage solutions like HBM and HBF, which are essential for GPU and AI server upgrades [1] - The investment approach for AI differs from that of the internet, with AI focusing on "to Task" models that enhance task efficiency through multi-model collaboration, contrasting with the internet's "to B" and "to C" models that prioritize user engagement and traffic [1] - The long-term investment strategies highlighted include trend investing, which capitalizes on short-term trends, and value investing, which focuses on the long-term growth of companies [1] Group 2 - For ordinary investors, a shift in mindset and investment methods is crucial, moving away from trend guessing and towards a value investment perspective that emphasizes long-term company growth [2] - A unique value investment formula proposed is: Value Investment ≈ Good Asset + Good Price, indicating that both components are essential for successful investment [2]

Core Insights - The storage module industry is experiencing a significant shortage driven by AI demand, with NAND Flash prices increasing by approximately 50% this month, and DRAM prices for DDR4 and DDR5 also rising [2] - NAND Flash applications in AI have surged recently, leading to a shift in supplier capacity towards AI applications, with expectations of a substantial market opportunity due to HDD shortages [2] - The major DRAM manufacturers have no plans to increase DDR4 production, anticipating a gradual exit from the DDR4 market by 2026, which limits supply and drives prices up [2] Group 1: Company Performance - In Q3, the company reported revenues of 4.109 billion, a quarter-over-quarter increase of 27.2% and a year-over-year increase of 63.2%, with a net profit of 1.511 billion, reflecting a quarter-over-quarter increase of 259% and a year-over-year increase of 334% [3] - The company is primarily focused on industrial control, with DDR4 still dominating product shipments, while expecting an increase in DDR5 shipments as customers transition [2] Group 2: Market Trends - SanDisk predicts that the NAND Flash supply shortage will persist until at least the end of 2026, with indications that the tight supply situation may extend into 2027 [4] - The demand for NAND Flash is driven by long-term trends, capital investments, and industry transitions, with data centers expected to become the largest segment for NAND Flash by 2026 [4] - SanDisk's wafer fabs are operating at full capacity to replenish significantly reduced inventories, and the company is optimistic about the long-term growth of the data center market [4][6] Group 3: Future Outlook - SanDisk's revenue guidance for Q2 FY2026 is projected between 2.55 billion and 2.65 billion, exceeding market expectations, with adjusted earnings per share forecasted between 3.00 and 3.40 [6] - The company reported a 26% quarter-over-quarter growth in data center revenue, with ongoing collaborations with major data center clients [7] - BiCS8 technology is expected to dominate production by the end of FY2026, currently accounting for 15% of total shipments [7]