EP57: Is China winning the Chip War

Summary:

In this episode, we examine China's position in the global chip race, particularly for advanced AI chips. We detail China's struggles to produce chips smaller than 7 nanometers (nm) due to export restrictions on crucial lithography machines and the difficulty of catching up with leading manufacturers. We then delve into the different types of AI chips and the companies developing them, highlighting that while China has several companies making AI chips, many rely on foreign technology or are lagging in terms of nm size, indicating a dependency on foreign manufacturers and a disadvantage in the race for cutting-edge AI technology.

Questions to consider as you read/listen:

1. What are the major factors hindering China's advancement in the global semiconductor and AI chip production

2. How do China's AI chip makers compare in terms of technology and production capacity to their global competitors?

3. What are the implications of China's current status in the chip war for the global technology landscape and future of artificial intelligence?

Long format:

Is China winning the chip war?

No.

Here are the facts:

1. China is producing in fabs (mainly at the Semiconductor Manufacturing International Corporation) 7nm chips but not at scale. They are doing nothing smaller than 7nm. The current state of the art is 3nm with Rapidus (Japan) saying it is over 80% complete making its 2nm fab that will produce at scale and then will move on to 1.4nm.

2. TSMC, Samsung and Japanese fabricators along with the US fabricator have agreed to export bans of chips of various sizes the smallest of which is 5nm and will likely be bans to anything under 7nm soon.

3. Semiconductor fabrication requires lithography machines to produce semiconductors. Above 7nm involves lithography technology called Deep Ultraviolet (DUV). DUV is made by ASML (Dutch), Nikon (Japan) and Cannon (Japan). China has no domestic DUV lithography machine production. To my knowledge with export licenses, DUVs can still be exported to China. In terms of lithography equipment, 90 nm-process lithography machines made by Shanghai Micro Electronics Equipment (SMEE) are the mature domestic product for mass production in China. The best China can do at present is the lithography machine for the 28 nm process.

4. To produce any chip under 7nm you need Extreme Ultraviolet (EUV) machines or you physically cannot “print” that small on a wafer. Only one company makes EUV. That is ASML of the Netherlands. The Dutch passed a law making it illegal to export EUV machines to China. ASML has agreed to comply. China has zero EUV machines in China.

5. Shanghai Micro Electronics Equipment Group (SMEE) is a Chinese state-owned enterprise that has filed a patent for EUV radiation generators and lithography equipment. However, experts say that SMEE still lags behind ASML and would need to overcome many restrictions to catch up.

6. Thanks to the Department of Commerce restrictions and agreement by the Dutch and Japanese, the parts and consumables and design for EUVs cannot be exported to China.

That’s what leads PZ to say that the Chinese can’t get EUVs (and therefore cannot make under 7nm where it is now), can’t import the finished sub 5nm chips (without breaches or workarounds which will always happen), and can’t get the “stuff” needed to make the EUVs it needs to get below 7nm.

The only real hope China has is that the Dutch put profits over politics as ASML took a huge valuation hit on October 15 and there is some talk in some quarters to rescind or alter the EIV export ban.

Is China winning the AI production battle?

What makes an AI chip?

AI Chips have parallel processing capabilities.

While general-purpose chips employ sequential processing, completing one calculation at a time, AI chips harness parallel processing, executing numerous calculations at once.

There are various different type of AI chips. Graphics processing units (GPUs), field programmable gate arrays (FPGAs) and application-specific integrated circuits (ASICs) are all considered AI chips.

GPUs

GPUs are designed to handle graphics and video rendering, and are often used for image classification, autonomous driving, and Safe City. GPUs are good for tasks that can be broken down into smaller components and completed in parallel.

FPGAs

FPGAs are programmable silicon chips that can be configured for specific applications. They are often used for deep learning and big data analysis. FPGAs are flexible and can be reprogrammed to optimize for specific applications, which can reduce power consumption and latency.

ASICs

ASICs are custom-made chips designed for a specific purpose. They are often used for data inference, AI synthesis, and assisted driving. ASICs are optimized for speed, power, and size, but they are expensive and time-consuming to design and manufacture. Once an ASIC is designed, its functionality is fixed

Here is a list of current AI chip makers: https://seo.ai/blog/ai-chip-makers

As we can see above, there are a number of Chinese companies. But here is recent detailed inventory of production.

Of the above chart note the node which is the nm sized chip and the fab.

The devil is in the details.

Huawei

A major player in the global chip design market, Huawei is working to develop AI chips locally. Huawei's Ass 910B chip was once considered the most competitive AI GPU designed in China. Huawei's second-generation AI chip, Ascend 920B,which was first seen in Chinese servers in 2022. It was made using TSMC's 7-nm process

SMIC

SMIC is one of China's best hopes for producing advanced AI chips, but it is still one to two generations behind in mature production. SMIC only began large-scale production of 14 nm chips in 2022, and its mature production remains with the 28 nm process, still one to two generations behind.

Baidu

Baidu designs its own chips for AI processes, including the Kunlun chip for servers and autonomous cars. They use Nvidia chips.

Alibaba

Alibaba's T-Head unit developed the Hanguang 800 AI inference chip, which is used to accelerate the recommendation system on its e-commerce platform. They use Nvidia chips.

Biren Technology

Biren Technology designs a general purpose GPU and has a software development platform to build applications on top of the hardware.

Objectively, China isn’t winning the chip war and is in a corner. And in terms of producing ADVANCED cutting edge AI chips, they are either totally dependent on foreign fabs or they are inferior because they are high nm size platforms or both.

Sources:

https://www.ibm.com/think/topics/ai-chip

https://cset.georgetown.edu/wp-content/uploads/AI-Chips%E2%80%94What-They-Are-and-Why-They-Matter-1.pdf

https://builtin.com/articles/ai-chip

https://forum.huawei.com/enterprise/en/differences-between-cpu-gpu-fpga-and-asic/thread/966389-895?page=1#:~:text=CPU:%20The%20CPU%20processes%20complex,assisted%20driving%2C%20and%20AI%20synthesis

https://www.ibm.com/think/topics/fpga-vs-gpu#:~:text=GPU%20challenges&text=FPGAs%20are%20programmable%20silicon%20chips,applications%20and%20prototyping%20new%20projects

https://www.wevolver.com/article/asic-vs-fpga#:~:text=Long%20Development%20Time:%20The%20ASIC,to%20update%20functionality%20is%20important

https://blog.purestorage.com/purely-educational/gpus-vs-fpgas-whats-the-difference/#:~:text=Choosing%20between%20GPUs%20and%20FPGAs,hardware%20might%20not%20be%20sufficient

https://www.linkedin.com/advice/0/what-key-differences-between-asic-fpga-soc#:~:text=6-,1%20ASIC:%20Application%2DSpecific%20Integrated%20Circuit,%2C%20graphics%20processing%2C%20or%20encryption

https://aws.amazon.com/compare/the-difference-between-gpus-cpus/#:~:text=The%20CPU%20handles%20all%20the,and%20finish%20them%20in%20parallel

https://www.cigionline.org/articles/in-the-global-ai-chips-race-china-is-playing-catch-up/#:~:text=Huawei%20and%20SMIC%20now%20stand,(fifth%2Dgeneration)%20chips

https://www.cnbc.com/2024/09/17/chinese-companies-aiming-to-compete-with-nvidia-on-ai-chips.html#:~:text=Alibaba%20and%20Baidu%20both%20buy,Cambricon%20Technologies