45 minute interview with Intel CEO Chen Liwu He mentioned in the interview why CPUs have been revived in the AI era? What is the difference between Intel's advanced packaging and TSMC's advanced packaging? There are also a bunch of materials that you often hear about, such as glass substrates, synthetic diamonds, gallium nitride, and phosphine. Why are they particularly important in the next 10 years? Chen Liwu explained it thoroughly through three fundamental shifts in his interview. ·The first one is to calculate the shift of the center of gravity. In the past three years, the market has equated the computing power of AI directly with GPU. Chen Liwu's judgment is very direct. The logic of the training era is valid, but the era of inference and intelligent agents is not. Because the intelligent agent is not constantly performing matrix multiplication, it needs to schedule tasks, check databases, manage memory, adjust tools, perform secure controls, and coordinate among hundreds of models, all of which are the work of the CPU. After chatting with the AI model manufacturer, he found that the ratio of GPU and CPU was increasing from 8:1 to 4:1 and then to 1:1. You don't have to kill Nvidia, the stronger the GPU sales, the greater the demand for CPU in the system, so Intel just needs to stay in its current position. ·The second one is the transfer of connection methods. The shrinking of transistors is becoming increasingly difficult and expensive, so the next stage of competition is not about who can make smaller nanoscale chips, but about who can assemble more chips more efficiently together. The approach of TSMC CoWOS is to lay a large silicon wafer underneath the GPU and HBM as a base, which is equivalent to repairing a whole elevated platform. The silicon area is limited and the yield is difficult to climb. Intel's EMIB has completely changed its approach, only connecting a small silicon bridge at the junction of two chips, and continuing to use cheap organic substrates elsewhere. One sentence analogy is that TSMC is building an elevated platform, while Intel only builds a bridge at the crossing of the river. Bridges are cheaper than platforms, faster, and have no limit on area. Combined with Foveros' 3D vertical stacking and copper hybrid bonding, the goal is to integrate the CPU, GPU, and HBM I/O into a complete system level package. ·The third one is the transfer of material systems, which is also the most information dense part of the entire interview. Chen Liwu clearly stated that the current trend is shifting towards new materials, and he proposed five key materials in one breath. ·Firstly, there is the glass substrate. Currently, all advanced packaging uses organic substrates, which can severely warp and deform under the ultra-high heat generated by kilowatt level AI chips. However, the flatness and thermal stability of glass substrates are 50% higher than those of organic materials, and they can create denser and finer through holes, making them the solid foundation for the next generation of giant AI chips. Chen Liwu adopted a pragmatic light asset model, shifting design standards from Intel to sourcing from Korean suppliers and using venture capital thinking to quickly mature the entire supply chain. Next is indium phosphide. As the cluster size of large models continues to grow, the delay and power consumption of traditional copper wire transmission become unbearable. Indium phosphide is the core material for manufacturing semiconductor lasers. Intel's deep layout of indium phosphide aims to accelerate the implementation of co packaged optics (CPO), directly packaging the photoelectric conversion module and the core computing chip together at the packaging level, achieving high-speed communication using light instead of electricity. Finally, it is necessary to solve the power supply and heat dissipation problems of the computing monster. An AI rack can easily generate tens of thousands or hundreds of thousands of watts, and traditional silicon-based power devices suffer from astonishing losses during voltage conversion. Chen Liwu proposed the layout of third-generation semiconductors such as gallium nitride and silicon carbide to maximize the compression of power supply losses in the last few centimeters before the current enters the chip. When extreme heat is generated in the dead corners of the core inside the chip, Intel introduces the material with the highest thermal conductivity in nature, artificial diamond, as a thermal diffusion layer in the packaging to instantly dissipate the heat and prevent the chip from reducing frequency due to overheating. Chen Liwu is not just talking about these materials, he has invested in some projects with real money and silver. The three transfers are just directions, but the direction is only the first step. Judging correctly does not mean landing. Chen Liwu created his own execution framework, crawling, walking, and running, to ensure that these three transfers would not become castles in the air. ·Crawling is to first enable the organization to fight. When he took over, Intel was a table driven company, reporting numbers layer by layer, waiting for approvals layer by layer, and there was no communication between departments. When he came up, he had all the engineering teams report directly to him, cutting off all the bureaucratic layers in the middle. The product line was large and refined, no longer making all kinds of chips, and the decision-making speed must return to the level of a startup company. If the organization was not adjusted, three transfers could not be pushed forward. ·The second one is to leave, which is to prepare enough ammunition for the three transfers. On the capital side, he did some textbook level things and even pulled Huang Renxun, who seemed to be a rival, to invest 5 billion US dollars in him. In 14 months, the book value turned into 25 billion US dollars. SoftBank has entered, and the US government has become its major shareholder, which is binding all key players into a community of shared interests. What you want from Nvidia is my OEM and packaging, what you want from SoftBank is AI infrastructure, and what the US government wants is local manufacturing capabilities. Everyone is playing at the same table. ·The third one is running, which means combining the results of the three transfers into a whole system. His goal is not to create a popular chip, but to integrate CPU, GPU custom chips, advanced processes, advanced packaging, and software ecosystem into a complete computing platform. Note that this is called a computing platform, not a chip he used to be good at. After watching Chen Liwu's interview, I feel that Intel's approach is actually beneficial for China. Because Intel's advanced packaging EMIB does not require the production capacity of expensive silicon intermediate layers, you can take the route of local silicon bridge and panel level packaging. So these packaging companies in China can follow up in this direction. China has a strong foundation in the glass substrate, diamond heat dissipation, gallium nitride, and other industrial chains. China should be a strong point in industrialization and engineering. So during these 45 minutes, Chen Liwu appeared to be talking about Intel, but in reality, he was talking about the AI era. The shift in semiconductor competition is the second half, and what the giants are really doing, the shift is an opportunity.