事件：当地时间3月2日，BIS将浪潮、龙芯等29家中国实体列入实体清单。其中浪潮、龙芯同时被列入脚注4实体（即涉及先进计算类芯片与超级计算机的实体），将限制其获取18类软件和技术。

我们看好外部形势趋紧之下，Chiplet技术方案由设计公司引领、先进封装赋能落地，从上游IP、EDA、设计到中游制造，再到下游封测，革新半导体产业链，重塑产业链价值，有望助力国产芯实现换道超车。看好封装公司估值处于历史相对低位，周期底部有望率先复苏，伴随2D封装到3D Chiplet发展，封装环节价值逐步提升。

Tightening macro environment opens prospects for chiplet technology solutions

The US Bureau of Industry and Security (BIS) added Inspur and Loongson among 29 China companies to its Entity List in early March. The two companies are also mentioned in footnote 4 (covering companies related to advanced computing chips and supercomputers), which restricts their access to 18 types of software and technologies. In the tightening macro environment, we see favorable prospects for chiplet technology solutions. Led by design companies and aided by advanced packaging traction, chiplets boost innovations in the semiconductor supply chain and help reshape the industrial value chain: from IP, EDA and design in the upstream market to manufacturing in the midstream, and packaging and testing in the downstream market. This could help core Chinese market players make the switch and help them overtake the competition. We believe that packaging companies, whose valuations are at historical lows, will bottom out cyclically. With developments from 2D packaging to 3D chiplets, we believe the value of packaging is on an uptrend. 

1、硅片级“解构-重构-复用”，Chiplet突破三大产业瓶颈

作为硅片级“解构-重构-复用”的方案，Chiplet或为破局摩尔定律、实现产业再度飞跃的关键。其突破主要体现在三个方面，分别带来百分比级、翻倍级和指数级的性能提升：（1）成本&良率突破：芯片良率与芯片性能之间存在矛盾，同构小芯粒集成可以提升良率，降低成本；（2）面积&性能突破：目前服务器CPU和GPU已逼近单个芯片面积上限，无法满足高算力需求，同构扩展可以提高性能，应对各场景大量增长的算力诉求；（3）设计&制程突破：先进制程芯片设计成本高昂，而异构集成下各模块使用适宜制程各司其职，从我国视角来看，Chiplet或为打破国产制程瓶颈的关键方案。同时，Chiplet可助力超异构集成计算的发展，处理器性能&灵活性同步提升。

3 wafer-level breakthroughs: deconfiguration, reconfiguration and multiplexing

As chiplets offer wafer-level deconfiguration, reconfiguration and multiplexing solutions, they are a key to breaking Moore’s Law and could bring another technological leap for the industry. The three key aspects are percentage, doubling and exponential performance improvements:

(1) cost and yield rates: in view of the conflict between chip yield rate and performance, the integration of isomorphic small chips could improve yield rate and reduce cost.

(2) area and performance: server CPUs and GPUs are approaching the upper limit of individual chip area and are unable to fulfill demand for high computing power. Isomorphic scaling could improve performance and meet massively expanding computing power demand in various scenarios.

(3) design and process: advanced process chip design is costly and the heterogeneous integration of each module uses an appropriate process to perform its duties. Chiplets could be a key solution to the process bottleneck in the Chinese industry. They can boost the development of ultra-heterogeneous integrated computing to improve processor performance and flexibility simultaneously.

2、产业链价值重塑，封装环节具有“估值处历史相对低位+周期复苏+产业价值量提升”的投资逻辑

从全产业链来看，Chiplet作为一种全新设计理念提升了设计、IP、EDA环节的引领性地位，有望为中游制造、下游封测带来价值增量。从具体的落地方案来看，Chiplet主要依靠高速互联的设计和异构集成先进封装技术的支撑。设计方面，主要通过Base Die/IO Die/Die to Die 设计实现核心处理模块之间，及其他各模块间的高速互联。封装方面，Chiplet封装演进的本质是在成本可控的情况下尽可能提升互联的密度与速度，从2D 封装到2.5D Chiplet、3D Chiplet，封装环节价值量&重要性有望不断提升。我国封装厂商技术积累深厚，长电科技、通富微电、华天科技已实现Chiplet量产，封装环节具有“估值处历史相对低位+周期复苏+产业价值量提升”的投资逻辑。

Packaging investment: low valuations; cyclical recovery; value chain upgrade

Chiplets are a new design concept that has impact across the entire industry chain. They have raised the leading positions of the design, IP and EDA segments. We expect they will keep adding incremental value to midstream manufacturing, and packaging and testing in the downstream. Their applicability includes the following: 

•Implementation solutions: chiplets rely on the support of high-speed interconnected design and heterogeneous integrated advanced packaging technology. 

•Design: high-speed interconnection between core processing modules and other modules is mainly realized through base die, IO die and die-to-die designs. 

•Packaging: the essence of the chiplet packaging evolution is to enhance the density and speed of interconnections as far as possible while keeping cost manageable. From 2D packaging to 2.5D chiplets and 3D chiplets, we expect the value and importance of the packaging business will keep increasing. China’s packaging manufacturers have built up deep technological know-how, with JCET Group, Tongfu Microelectronics and Huatian Technology achieving chiplet mass production. Our investment rationale for the packaging industry rests ties three aspects: historically low valuations, a potential cyclical recovery and value chain upgrade.

3、高性能计算（HPC）或为Chiplet当前的主要发力点

ChatGPT是大数据+大模型+大算力的产物，每一代GPT模型的参数量高速增长，根据人工智能学家公众号数据，2020年5月发布的ChatGPT的前身GPT-3参数量达到了1750亿（预训练数据量达45TB，远远大于GPT 2的40GB）。算力需求方面，训练ChatGPT所耗费的算力大概是3640 PetaFLOPs per day，即用每秒能够运算一千万亿次的算力对模型进行训练，需要3640天完成。随着科技巨头类ChatGPT项目入局，整体在算力提升、数据存储及数据传输端需求迭起。而随着摩尔定律逐渐趋缓，我们认为Chiplet有望成为支持高性能计算存储的关键。美国正在开发的三个超级计算机Aurora、El Capitan和Frontier，CPU和GPU利用Chiplet方案，AMD，Intel，华为的服务器处理器芯片均采用Chiplet方案助力算力突破及性能提升。

Chiplets help to drive high-performance computing requirements

ChatGPT is the product of big data, the large language model and high computing power. The parameters of each generation of GPT model are rising rapidly. According to the public accounts of artificial intelligence scientists, GPT-3, the predecessor of ChatGPT released in May 2020, has a parameter volume of 175bn (with pre-training data at 45TB, much larger than GPT-2’s 40GB). In terms of computational power requirements, 3,640 PetaFLOPs per day was expended to train ChatGPT. In other words, it takes 3,640 days to train the model with computational power that is capable of performing 1 quadrillion calculations per second. With the entry of technology giants’ ChatGPT projects, demand for computing power enhancement, data storage and data transmission iteration continues to rise. As Moore’s Law slows, we believe chiplets will provide the key support for high-performance computing storage. The three supercomputers under development in the US – Aurora, El Capitan and Frontier – use chiplet solutions for CPUs and GPUs. AMD, Intel and Huawei’s server processor chips all use chiplet solutions to improve computing power and performance.