Track the latest AI trends

AI reshapes the PC industry chain: Microsoft opens the battlefield. Qualcomm, Intel, and AMD scuffle CPUs, and memory benefit the most

wallstreetcn · May 9 13:08

来源：华尔街见闻作者：赵雨荷

摩根士丹利研报称，要在PC设备上正确运行Copilot，人工智能PC需要配备每秒45万亿次操作（TOPS）的计算能力，而这将驱动CPU和内存的升级。其中内存将成为关键，预计32GB内存将成为人工智能PC的默认配置，人工智能PC不断提高的渗透率也将加剧内存的供不应求。

人工智能正在改变PC领域，使其变得更智能、更快、更加个性化。摩根士丹利近日发布研报称，未来的PC市场将由人工智能PC主导，而这将驱动CPU和内存的升级。其中内存将成为关键，预计32GB内存将成为人工智能PC的默认配置，人工智能PC不断提高的渗透率也将加剧内存的供不应求。

人工智能PC的变化来自哪里：满足微软Copilot运行需要的45 Tops算力

研报认为，目前已有的PC大多已经老化，人工智能则可以改变未来的工作方式。摩根士丹利将人工智能PC定义为具备特定SoC（片上系统）功能的超级个人计算机，能够在本地运行生成式AI任务。由于目前大多数CPU或GPU并未针对高效运行AI任务进行优化，而人工智能PC配备的神经处理单元（NPU）将能够高效运行这些任务。到目前为止，仅存在配备低于40 TOPS（每秒数万亿次操作）NPU的PC芯片，例如英特尔的Core Ultra系列、AMD的Ryzen 7040系列、高通的Snapdragon 8cx系列和苹果的M系列。

目前微软正努力在最新版本的Windows中提供Copilot和其他AI功能，第三方开发者随后将创建相关应用程序。而要在PC设备上正确运行Copilot，人工智能PC需要配备每秒45万亿次操作（TOPS）的计算能力。即将推出的AI功能可能会更依赖设备硬件，这意味着微软将推动其制造合作伙伴，在新的笔记本和台式机的最低内存配置设定为16GB，以包含Copilot和其他Windows 11的AI功能。升级内存是提高PC性能最简单的方法之一，这将是自2012年后PC时代以来在PC内存行业看到的首次重大升级。

随着Windows 12的推出，摩根士丹利的全球PC硬件团队预计，AI笔记本电脑的普及率将在2025年迅速增长。研报估计2025年和2026年人工智能PC的渗透率将分别达到30%和50%（从2024年的8%提升）。人工智能PC不仅为戴尔等系统制造商带来了更高的利润率，还将每台PC的半导体含量提高了20%-30%，主要来自CPU和内存的升级（这为全球半导体行业增加了约300亿美元的收入）。

CPU大战：X86架构 VS Arm架构

支持AI的PC处理器种类繁多。$高通 (QCOM.US)$的Snapdragon X Elite平台是第一个符合微软Copilot标准的产品，预计将于2024年下半年开始发货，并提供约45 TOPS的计算能力，超过微软定义的40 TOPS的人工智能PC要求。AMD的Ryzen 8000系列（Strix Point）与英特尔的Meteor Lake一起于2023年12月发布，具备34 TOPS的CPU+GPU+NPU综合计算能力，尚未达到微软的标准，不过即将推出的Lunar Lake处理器将在今年年底前突破40 TOPS的门槛。

研报表示，英特尔和AMD的x86 CPU架构与高通ARM架构的CPU在人工智能PC市场的竞争将加剧。高通率先符合微软的要求，使其能够抓住第一波人工智能PC的机会，像戴尔、$惠普 (HPQ.US)$、联想、华硕和宏碁这样的主要PC原始设备制造商都将在2024年推出搭载高通CPU的机型，挑战英特尔的x86架构阵营。

内存收益逻辑

研报认为，内存在人工智能PC的整体系统性能和AI任务执行效率中起着至关重要的作用。AI操作需要大量数据进行训练和深度学习，这些数据会临时存储在内存中。经过训练的AI模型包含大量参数，随着模型复杂性的增加，参数数量也相应增加，因此需要足够的内存容量来存储这些参数。因此，高速大容量内存可以提供更快的数据访问速度，并有助于加速模型训练过程。

根据研报，内存在x86与ARM的竞争中受益匪浅。人工智能PC推动了平均PC内存容量的显著增长，并且在PC内存中使用了更高的LPDDR。AI加速是一种高度依赖内存的操作，大型语言模型（LLM）需要大量的快速频繁访问的内存。这就是微软将人工智能PC中的内存基准容量设定为16GB的原因，但预计大多数客户和企业可能会选择32GB为默认配置，以获得更好的性能并做好未来准备。

16GB不仅用于本地任务加速，还可用于云端的Copilot AI功能。这一容量比2023年平均每台PC的内存容量高出50%，当时约为10.5GB。摩根士丹利预测，人工智能PC将推动对PC内存需求增长更快，而消费者升级趋势可能进一步提升这一需求增长轨迹。

此外，大多数人工智能PC的CPU将采用LPDDR5x，而不是目前主流的DDR SO-DIMM模块。带宽和能源效率对人工智能PC来说同样重要。这种低功耗内存的转变是由更快数据传输的需求驱动的——DDR5的速度在4.8–5.6 Gbps之间，而LPDDR5x达到7.5–8.5 Gbps，能够满足人工智能PC快速语言处理和响应的需求。与DDR5相比，LPDDR5X在主动使用时的功耗降低了约50%，在自刷新时的功耗降低了86%，并且带宽更高，达到了6.4 Gb/s，而DDR5的带宽为4.8 Gb/s。摩根士丹利估计今年LPDDR将占PC内存需求的30–35%，未来的增长将由人工智能PC的更高普及率和CPU制造商的持续采用下推动。

而人工智能PC的普及和渗透率增加对整体更新周期最终是有利的，特别是在商业领域。摩根士丹利的大中华区硬件团队预测，到2027年底，人工智能PC的渗透率将达到53%，意味着人工智能PC内存将在2027年占全球内存出货量的10%。尽管2024年人工智能PC内存仅占全球内存市场的3%，但其内存容量的提升仍将加剧已经紧张的内存供应，摩根士丹利估计PC内存的供需比将从目前的2%下降到-9%，全球内存的总供需比将下降2个百分点至-5%。

杀手级应用会来自哪里

摩根士丹利认为，目前微软的Copilot是主要的杀手级AI应用，而随着渗透率的提高，第三方开发者也将会加入。

Copilot是一款由AI驱动的生产力工具，它协调大型语言模型和$微软 (MSFT.US)$ Graph中的内容，帮助用户日常操作，所有用户都可以通过Copilot网站、必应聊天、Microsoft Edge中的Copilot和Windows中的Copilot访问。Copilot还可以通过Copilot、必应、Edge、Microsoft Start和Microsoft 365移动应用程序访问。

同时，微软还将Copilot嵌入到主要应用程序中，例如Word、Excel、PowerPoint、OneNote、Teams和Outlook。例如，PowerPoint中的Copilot可以将现有的Word文档转换为演示文稿；Excel中的Copilot将帮助分析和探索数据，以帮助做出最佳决策，并生成图表来直观展示机会；Word中的Copilot将撰写、编辑、总结、提供建议并协同创建。此外，当在Teams上进行对话/视频会议时，它还可以帮助生成会议纪要。

目前，Copilot为个人用户提供免费和专业版，其中专业版允许用户在高峰时段优先访问GPT-4和GPT-4 Turbo以获得更快的性能，构建定制的Copilot GPT，以满足个人需求和兴趣。尽管目前Copilot在云端运行，但摩根士丹利预计，在进入Windows 12时代时，人工智能PC将推出离线版，这将有助于缓解有关数据安全和可能的网络速度下降的担忧。

编辑/lambor

Source: Wall Street News Author: Zhao Yuhe

According to the Morgan Stanley Research Report, in order to properly run CoPilot on PC devices, artificial intelligence PCs need to be equipped with 45 trillion operations per second (TOPS) of computing power, which will drive CPU and memory upgrades. Among them, memory will be the key. It is expected that 32GB memory will become the default configuration for artificial intelligence PCs, and the increasing penetration rate of artificial intelligence PCs will also increase the shortage of memory supply.

Artificial intelligence is transforming the PC sector, making it smarter, faster, and more personalized. Morgan Stanley recently released a research report stating that the future PC market will be dominated by artificial intelligence PCs, which will drive CPU and memory upgrades. Among them, memory will be the key. It is expected that 32GB memory will become the default configuration for artificial intelligence PCs, and the increasing penetration rate of artificial intelligence PCs will also increase the shortage of memory supply.

Where did the changes in artificial intelligence PCs come from: 45 Tops computing power to meet the operating needs of Microsoft Copilot

According to the research report, most existing PCs are aging, and artificial intelligence can change the way we work in the future. Morgan Stanley defines an artificial intelligence PC as a superpersonal computer with specific SoC (system-on-chip) functions, capable of running generative AI tasks locally. Since most CPUs or GPUs are currently not optimized to run AI tasks efficiently, artificial intelligence PCs equipped with neural processing units (NPUs) will be able to run these tasks efficiently. Until now, there are only PC chips equipped with less than 40 TOPS (trillions of operations per second) NPU, such as Intel's Core Ultra series, AMD's Ryzen 7040 series, Qualcomm's Snapdragon 8CX series, and Apple's M series.

Currently, Microsoft is working to provide Copilot and other AI features in the latest version of Windows, and third-party developers will later create related applications. To properly run CoPilot on a PC device, an artificial intelligence PC needs to be equipped with 45 trillion operations per second (TOPS) of computing power. Upcoming AI features are likely to be more dependent on device hardware, which means Microsoft will push its manufacturing partners to set a minimum memory configuration of 16GB in new notebooks and desktops to include Copilot and other Windows 11 AI features. Upgrading memory is one of the easiest ways to improve PC performance, and this will be the first major upgrade seen in the PC memory industry since the post-2012 PC era.

With the launch of Windows 12, Morgan Stanley's global PC hardware team anticipates rapid growth in the penetration rate of AI laptops in 2025. The research report estimates that AI PC penetration rates will reach 30% and 50% in 2025 and 2026, respectively (up from 8% in 2024). Artificial intelligence PCs not only brought higher profit margins to system manufacturers such as Dell, but also increased the semiconductor content of each PC by 20%-30%, mainly due to CPU and memory upgrades (which increased the global semiconductor industry's revenue by about 30 billion US dollars).

CPU Wars: X86 architecture vs. Arm architecture

There are many types of AI-enabled PC processors.$Qualcomm (QCOM.US)$The Snapdragon X Elite platform is the first product to meet Microsoft's Copilot standards. It is expected to be shipped in the second half of 2024, and will provide approximately 45 TOPS of computing power, exceeding the 40 TOPS artificial intelligence PC requirement defined by Microsoft. AMD's Ryzen 8000 series (Strix Point), along with Intel's Meteor Lake, was released in December 2023. It has 34 TOPS of CPU+GPU+NPU comprehensive computing power and has not yet reached Microsoft's standards, but the upcoming Lunar Lake processor will break the 40 TOPS threshold by the end of this year.

According to the research report, competition between Intel and AMD's x86 CPU architectures and Qualcomm ARM CPU architectures will intensify in the artificial intelligence PC market. Qualcomm was the first to meet Microsoft's requirements, enabling it to seize the opportunities of the first wave of artificial intelligence PCs, such as Dell,$HP Inc (HPQ.US)$Major PC original equipment manufacturers such as Lenovo, Asus, and Acer will all launch models equipped with Qualcomm CPUs in 2024, challenging Intel's x86 architecture camp.

Memory yield logic

According to the research report, internal memory plays a critical role in the overall system performance of AI PCs and the efficiency of AI task execution. AI operations require large amounts of data for training and deep learning, and this data is temporarily stored in memory. The trained AI model contains a large number of parameters, and as the complexity of the model increases, the number of parameters also increases accordingly, so sufficient memory capacity is required to store these parameters. As a result, high-speed, large-capacity memory can provide faster data access and help speed up the model training process.

According to the research report, internal memory has greatly benefited from the competition between x86 and ARM. Artificial intelligence PCs have driven a significant increase in average PC memory capacity and use higher LPDDR in PC memory. AI acceleration is a highly memory-dependent operation, and large language models (LLM) require large amounts of fast and frequently accessed memory. That's why Microsoft has set the base amount of memory in AI PCs at 16GB, but it's expected that most customers and businesses will probably choose 32GB as the default configuration to get better performance and be prepared for the future.

16GB is used not only for local task acceleration, but also for Copilot AI functions in the cloud. This capacity is 50% higher than the average memory capacity of each PC in 2023, when it was around 10.5GB. Morgan Stanley predicts that artificial intelligence PCs will drive demand for PC memory to grow faster, and consumer upgrading trends may further enhance this demand growth trajectory.

Furthermore, most AI PC CPUs will use LPDDR5x instead of the current mainstream DDR SO-DIMM module. Bandwidth and energy efficiency are just as important for AI PCs. This shift in low-power memory is driven by the need for faster data transmission — DDR5 speeds are between 4.8-5.6 Gbps, while LPDDR5x reaches 7.5-8.5 Gbps, which can meet the needs of AI PCs for fast language processing and response. Compared with DDR5, the power consumption of LPDDR5X is reduced by about 50% during active use, the power consumption during self-refresh is reduced by 86%, and the bandwidth is higher, reaching 6.4 Gb/s, while the bandwidth of DDR5 is 4.8 Gb/s. Morgan Stanley estimates that LPDDR will account for 30-35% of PC memory requirements this year, and future growth will be driven by the higher penetration rate of artificial intelligence PCs and the continued adoption of CPU manufacturers.

However, the increase in the popularity and penetration rate of artificial intelligence PCs is ultimately beneficial to the overall update cycle, especially in the commercial sector. Morgan Stanley's Greater China hardware team predicts that by the end of 2027, the penetration rate of AI PCs will reach 53%, which means AI PC memory will account for 10% of global memory shipments in 2027. Although artificial intelligence PC memory accounts for only 3% of the global memory market in 2024, the increase in memory capacity will still exacerbate the already tight memory supply. Morgan Stanley estimates that the PC memory supply and demand ratio will drop from the current 2% to -9%, and the total global memory supply and demand ratio will drop by 2 percentage points to -5%.

Where do killer apps come from

Morgan Stanley believes that currently Microsoft's Copilot is the main killer AI application, and as penetration increases, third-party developers will also join.

Copilot is an AI-driven productivity tool that coordinates large language models and$Microsoft (MSFT.US)$ The content in Graph helps users operate on a daily basis. All users can access it through the Copilot website, Bing Chat, Copilot in Microsoft Edge, and Copilot in Windows. Copilot is also accessible via the Copilot, Bing, Edge, Microsoft Start, and Microsoft 365 mobile apps.

At the same time, Microsoft also embeds Copilot into major applications such as Word, Excel, PowerPoint, OneNote, Teams, and Outlook. For example, Copilot in PowerPoint can turn an existing Word document into a presentation; Copilot in Excel will help analyze and explore data to help make the best decisions and generate diagrams to visualize opportunities; and Copilot in Word will write, edit, summarize, provide suggestions, and create collaboratively. Additionally, it can help generate meeting minutes when conversations/video meetings are taking place in Teams.

Currently, Copilot offers free and professional versions for individual users, with the professional version allowing users to prioritize access to GPT-4 and GPT-4 Turbo during peak hours for faster performance, and to build customized Copilot GPT to meet individual needs and interests. Although Copilot currently runs in the cloud, Morgan Stanley anticipates that AI PCs will launch offline versions when entering the Windows 12 era, which will help ease concerns about data security and possible network speed drops.

edit/lambor

The translation is provided by third-party software.

The above content is for informational or educational purposes only and does not constitute any investment advice related to Futu. Although we strive to ensure the truthfulness, accuracy, and originality of all such content, we cannot guarantee it.