更多资料加入知识星球:水木调研纪要 关注公众号:水木纪委 Strategy Latest in Al: Antitrust inquiries; chips at Computex; Tesla, OpenAl leadership David Folkerts-Landau, Ph.D. Group Chief Economist Global Head of Research +44-20-754-55502 Global Date 11 June 2024 In this edition of 'Latest in AI' we also discuss the importance of the rapid pace of chip unveilings from Nvidia, AMD and Intel at the annual Computex trade show in Taiwan. Furthermore, we analyse the corporate leadership issues sparked by Elon Musk at X and Tesla, and OpenAI' ...

Q：为什么超算需要使用高功率密度的冷却方式，而不能像普通计 育那样通过增大房间或者增加机器来解决？ | --- | --- | |-------------------------------------------------------------|-------| | | | | A：超算的结构决定了它需要使用高功率密度的冷却方式。因为超 | | | 算是一堆 CPU 的堆叠，CPU 的热度是相对均匀的，所以泡在新工 | | | 业里面这种方案是比较好的一个方案。这样可以在更小的面积之内 | | | 贴更多芯片，推出来的算力会更高。而对于 GPU 这种发热不均匀的 | | | | | 设备，采用冷板这种方式效果最好。 Q：如果采用降变式液冷，成本会有多大的增加? A：如果采用降变式液冷，成本会在普通相变的基础上增加 50%到 100%。这是因为静默式相变是一个成本较高的方案，所以在商用化 的数据中心领域中，基本上都是国家超算会使用。 Q：在什么情况下会使用静默式液冷，以及对未来这个产业的发展 有何看法? 冷系统的主要成本构成包括 CDU、many fold、二次管路和冷却塔 等。 Q&A Q：产业界对于 ...

谢谢好的各位领导晚上好然后我是华步电子训学然后给各位领导就是汇报一下就是我们针对5月份的话台股各个产业链的一个最新的一个营收包括紧急度展望的一个情况然后我们核心的结论就是说5月份的话这里上需求的一个还是在一个稳步复苏的状态里面然后这里面的话 像包括存储的话呢有持续改善然后像消费电子里面的话呢可以期待后面的一个APC对于整个PC大厂的一个拉动那么各个系统板块里面的话呢这个这个增长最明确的肯定还是整个AI产业链啊然后整个AI产业链的不管是代工然后像夜冷啊等等的各个公司其实五月份的营收环境都还都还很强劲 那么像其他的像元器件和这些功率器件的话整体上呈现一个稳步复苏的一个状态整体上我们认为就整个半导体还是去保持一个复苏的一个趋势那么其中AI还是一个非常强劲的一个拉动 那么接下来我们就去具体去分析一下各个环节的一个具体表现情况然后这个是五月份的一个各个各个板块的各个公司的一个营收环比的一个表现啊可以看到就是说呃像代工环节的话呢其实环比就是比较比较稳定啊然后在下游的IT设计其实环境表现还是可圈可点的像其中的话像SoC厂商环比有个小个增长 那像模拟包括说DDIC其实华文比的增长也还可以那特别是像MCU这块其实我们可以提醒一 ...

本次电话会议仅服务于长江证券研究所白名单客户未经长江证券事先书面许可任何机构或个人不得以任何形式对外公布、复制、刊载、转载、转发、引用本次会议相关内容否则由此造成的一切后果及法律责任由该机构或个人承担长江证券保留追究其法律责任的权利 尊敬的各位投资者大家晚上好我是长江证券基层第一行业贝鲁鲁中非常感谢各位领导接受到我们今天晚上的专题汇报我们今天的汇报主要是就最近苹果的W2WTC大会做一个从基层第一行业需要的解读 那首先的话呢就是这个事件背景是在于说呃应该是6月11号晚上明天1点苹果的这个2024年全球开幕招待会啊这是开幕啊在这个会里面啊从我们学生机行业视角来看最重要的就是苹果是发布了他的这个Apple Intelligence他的这个苹果智能啊嗯这块的话呢就是从市场的气概来看啊他应该是 终于带来了预期比较久的最新人工智能战略 因为过往在其他的手机或者是包括说之前的微软在这种电脑上面都是实现了AI的一些升级苹果这边的话大家一直是期待比较高所以它的AI进展一直受到社团关注从出来以后 在昨天晚间 今天凌晨的时候苹果股价应该是涨了大概7.26左右 创下了一个新高市值是达到了3.18万亿 也是获得了市场认可 从整个的发布 ...

Hello everyone, good afternoon. Welcome to Xinhua's live broadcast room. XIN, can you check if our sound and screen are normal? Let's check if our sound and screen are normal. Yes, I just checked if our sound and screen are normal. Today, we are very happy to invite our industry researcher, Mr. He Ye Lin, to our Xinhua S1 live stream. Welcome, Mr. He Ye Lin. Mr. He Ye Lin, please say hello to our viewers in the live stream. Good afternoon, everyone in the live stream. Thank you for participating in our shar ...

Summary of Key Points from the Conference Call Industry Overview - The conference call discusses the impact of artificial intelligence (AI) on storage and thermal management technologies, particularly in mobile devices and PCs, highlighting significant challenges and opportunities in these areas [2][3][4]. Core Insights and Arguments 1. **Increased Demand for High-Performance Memory** AI's growth has led to a significant increase in demand for high-performance memory and flash storage, with memory currently being the main bottleneck for AI operations. For instance, a 7 billion parameter model requires approximately 4GB of memory, while Apple emphasizes the need for devices supporting AI to have at least 8GB of LPDDR5 memory and sufficient bandwidth [2][3][7]. 2. **Storage Capacity and Speed Requirements** The deployment of large models on end devices has created unprecedented demands for storage capacity and speed. Despite advancements in model quantization technology, which reduces model size and increases inference speed, the rapid growth in data demand remains unmet, necessitating significant upgrades in hardware configurations for mobile and PC devices [4][9]. 3. **Challenges in Storage Bandwidth** The increasing requirements for memory capacity and data transfer rates pose major challenges for storage bandwidth. Traditional von Neumann architecture leads to frequent data transfers, consuming substantial time and power. The industry is exploring solutions like processing-in-memory and Memory on Package technologies to reduce data movement overhead and enhance performance while lowering power consumption [5][6][11]. 4. **Thermal Management Issues** As memory sizes and transfer rates increase, the proximity of storage to processing units exacerbates thermal challenges. Optimizing AI performance will require enhanced motherboard designs and thermal management capabilities, considering power reduction and efficient cooling technologies to support AI workloads [6][12]. 5. **Future Growth in Memory Capacity** It is anticipated that AI model parameters will significantly increase in the next one to two years, potentially doubling memory capacity requirements for mobile devices. This is driven by the need to run multiple applications simultaneously and to handle vast amounts of data for AI tasks [8][10]. 6. **Impact on Storage Market** The demand for memory and flash storage in mobile and PC markets is expected to rise dramatically due to AI. DRAM accounts for approximately 35% of memory usage in mobile and PC devices, with a combined share of about 51%. The server market, as a core downstream segment, will also benefit from increased storage demand driven by AI training and cloud inference [10]. 7. **Advancements in Packaging Technology** AI workloads necessitate high-speed data exchanges between storage units and processors, highlighting the importance of improving transfer rates and reducing energy consumption. Next-generation trends may include in-memory computing to eliminate data transfer boundaries, thereby lowering costs and power usage. Advanced packaging technologies are expected to facilitate this by concentrating data closer to computing units, significantly reducing data movement delays and power consumption [11][12]. Other Important Insights - The overall storage market is projected to thrive, particularly in the mobile and PC sectors, as AI applications proliferate and drive hardware upgrades [4][9]. - The integration of multiple localized large models in devices, such as Microsoft's Copilot PC, requires substantial hard disk space, with total installation sizes reaching approximately 35GB and operational space needs potentially nearing 100GB [7].

Financial Data and Key Metrics Changes - The overall impression from the North American AI industry research indicates that the progress of large models has plateaued compared to the previous year's excitement surrounding ChatGPT, while the demand for computing power remains strong due to a clear supply-demand tightness [9][11]. Business Line Data and Key Metrics Changes - The application of AI in products is facing challenges, with a notable gap in user expectations versus the capabilities of current large models, particularly in creative processes like PPT and Excel, which are complex and prone to cumulative errors [21][22]. Market Data and Key Metrics Changes - The competitive landscape shows that traditional tech giants like Google and Meta are still leading in talent acquisition and investment in R&D, although the gap with OpenAI is narrowing as the industry matures [20][15]. Company Strategy and Development Direction - The focus is shifting towards B-end applications that address current needs and improve efficiency, rather than creating new demands. Innovations in system-level terminals like AI PCs and AI phones are seen as crucial for C-end applications [22][19]. Management Comments on Operating Environment and Future Outlook - Experts express cautious optimism regarding the continuous iteration of large models over the next 2-3 years, although the pace of progress is not as rapid as before. The sustainability of scaling laws and the effectiveness of current model architectures remain under scrutiny [11][12][13]. Other Important Information - The research highlights that the current bottleneck in AI applications is primarily in the C-end native products, with a stronger focus on B-end demands. There is a need for adjustments from hardware manufacturers to unlock innovation [23][22]. Q&A Session Summary Question: Is the capability of large models still progressing? - Experts maintain a cautiously optimistic view on the iterative progress of large models over the next 2-3 years, although this optimism is based on specific advancements rather than a broad expectation of revolutionary changes [11]. Question: What are the current challenges in AI applications? - The main challenges include high user expectations and the complexity of creative processes, which lead to significant cumulative errors in applications like PPT and Excel [21][22].

更多资料加入知识星球：水木调研纪要 关注公众号：水木纪要 美国 AI 产业调研内容及纪要 主 1、要 大对 模象 型的迭代 加 V ： s h u i n u 9 8 7 0 据 数 模型：OpenAI、Anthropic、Amazon科学家、Meta科学家 研 报 和 要 纪 生态：NV硬件专家、NV软件工程师，涉及底层软硬件、开发平台及工具研 调 手 一 多 更 2 微、 软大 A模 IC型 R应 M、用 Adobefirefly 加 V ： s h u i n u 9 8 7 0 据 数 TeslaFSD 报 研 和 要 纪 研 调 手 一 3、算力产多业链 更 ...

更多一手调研纪要和研报数据加V：shuinu9870 国内模型研发的差距，从最前沿的科学角度差的依然比较大，硅谷的环境非常与世隔绝， 科技大厂对于研发的投入更纯粹（meta，amazon 的专家表示公司烧百亿美金，也并没有要 求一定药做出啥产品来），因此在前沿架构的探索上，和硅谷差距非常大。但是考虑到当前 硅谷整体模型的迭代也出现一定程度放缓，国内模型应该能够保持不被拉开。 4. 应用创新：北美在创新一样迷茫，瓶颈何在？ 更多一手调研纪要和研报数据加V：shuinu9870 总体上，对于应用的拓展，类似移动互联网时代的原生 AI 应用的，相比于创造需求， 解决当下需求，提高效率应该是当前 AI 能够看到机会，因此从 B 端角度入手，会是当前 AI 应用的主力方向。此外以 AIpc 和 AIphone 这种系统级别的终端创新，打破单个 APP 的数据 割裂，会是 C 端应用最重要的尝试方向（anthropic 专家观点）。 更多资料加入知识星球：水木调研纪要 关注公众号：水木纪要 小结： 更多一手调研纪要和研报数据加V：shuinu9870 本次调研主要走访硅谷 AI 相关公司，总体印象是，AI 的大模型和算力 ...

Financial Data and Key Metrics Changes - The AI industry in North America is still in the exploratory phase, with significant investments in talent and funding [1] - The overall growth in AI API usage is rapid, particularly in internal search and marketing, although customer willingness to pay is increasing slowly [14] Business Line Data and Key Metrics Changes - In the 2B application sector, companies like Adobe and Microsoft are integrating AI into existing software, but the growth in paid usage is not yet substantial [14] - For 2C applications, ChatGPT faces limitations due to a lack of essential users and legal concerns, while Google is focusing on developing multimodal applications [15] Market Data and Key Metrics Changes - The competitive landscape shows that domestic AI talent is on par with international standards, with the main differences being talent density and computational resources [11] - The performance of Nvidia's hardware is significantly ahead, while Google's TPU ecosystem is evolving rapidly [13] Company Strategy and Development Direction - Nvidia is focusing on building a software ecosystem to reduce client development and operational costs, while not being overly concerned about competition from AMD and Intel [7] - The industry is expected to see a shift towards multimodal AI, as the progress in language models may face bottlenecks [18] Management Comments on Operating Environment and Future Outlook - The management acknowledges that the AI application landscape in North America has not yet found significant breakthroughs, raising concerns about future capital expenditures [18] - There is an expectation that the technology gap between China and the US in AI will not widen significantly, but it will be challenging to close in the short term [13] Other Important Information - The salary for top AI talent in companies like OpenAI and Google is significantly higher than in traditional sectors, indicating a competitive hiring environment [9] - The maturity of robotics and autonomous driving technologies is still uncertain, with expectations for significant advancements in the next 3-5 years [15] Q&A Session Summary Question: What are the key challenges facing AI applications in North America? - The current application scenarios are similar to those in China, with no rapid growth or large-scale opportunities observed, leading to skepticism about future capital investments [18] Question: How does the talent landscape compare between domestic and international players? - Domestic AI talent is considered competitive, with no significant widening of the gap, although challenges remain in attracting talent from North America [16]