Chinese firms react to TSMC chip supplies halt and escalating US tension

[Daniel Nenni]

Credit: AFP

TSMC has halted providing advanced AI chips at 7 nanometers and below to Chinese customers, highlighting mounting challenges for China's semiconductor industry. The suspension follows the discovery of TSMC-produced chips in Huawei devices and comes as the incoming Trump administration signals potential stricter sanctions. At the IC China 2024 in Beijing, industry leaders expressed both caution and resolve, calling for intensified efforts in generative AI and semiconductor technology to capitalize on China's vast market potential.

The industry gathering brought together over 500 companies, including Naura, YMTC, CXMT, and Huawei, showcasing China's capabilities in integrated circuit design, foundry services, and packaging. Industry leaders noted that escalating Sino-US geopolitical tensions have become a key concern, with businesses bracing for further disruptions. Shenzhen Basic Semiconductor CEO Weiwei He revealed the company has invested an additional US$2.8 million to US$4.1 million to localize manufacturing and materials in China, mitigating risks from reliance on US and Taiwanese suppliers.

The intensifying US-China trade conflict has already strained Chinese startups. He recounted how supply chain uncertainties during the 2018 trade war forced a strategic pivot, including fears of losing access to Taiwanese manufacturers. These pressures have slowed expansion plans, particularly as American automotive clients explicitly reject Chinese-made products.

Amid these challenges, Moore Threads co-founder Longfei Dong expressed optimism, emphasizing that GPUs, unlike smartphone chips, do not require cutting-edge nodes and can leverage alternative manufacturing techniques. He called for collaboration between government, industry, and academia to tackle production bottlenecks. However, Moore Threads continues to be hamstrung by its inclusion on the US trade blacklist, limiting its ability to rival Nvidia.

TSMC's suspension of services for Chinese AI chipmakers represents another blow, even as strong demand from smart hardware manufacturers in Shenzhen is expected to bolster TSMC's 2024 sales. Industry analysts note that while US-led decoupling efforts hinder China's access to advanced technology, the country continues to spend more on chips than any other nation.

China's ambitions are strengthened by Beijing's latest economic stimulus measures, offering growth opportunities for local players like SMIC and Hua Hong. Yet, restrictions on acquiring advanced wafer manufacturing equipment leave China struggling to compete in AI and other cutting-edge fields. Huawei's efforts to produce high-performance chips remain stymied, highlighting the deepening technological divide between China and the West.

Chinese firms react to TSMC chip supplies halt and escalating US tension

TSMC has halted providing advanced AI chips at 7 nanometers and below to Chinese customers, highlighting mounting challenges for China's semiconductor industry. The suspension follows the discovery of TSMC-produced chips in Huawei devices and comes as the incoming Trump administration signals...

www.digitimes.com

 

[KevinK]

US$2.8 million to US$4.1 million to localize manufacturing and materials in China, mitigating risks from reliance on US and Taiwanese suppliers.

Seems like a drop in the bucket compared to how much money would truly be required for long term mitigation.

Moore Threads co-founder Longfei Dong expressed optimism, emphasizing that GPUs, unlike smartphone chips, do not require cutting-edge nodes and can leverage alternative manufacturing techniques.

Boy is he wrong, unless you are talking about edge-based inference for smallish models. But I guess he has to spout unicorns and rainbows.

 

[Daniel Nenni]

Boy is he wrong, unless you are talking about edge-based inference for smallish models. But I guess he has to spout unicorns and rainbows.

Agreed. Quite the opposite I would say. Do smartphones really need the latest and greatest processor technology right out of the gate? If you have bloated software maybe. Do you really need a terabyte of storage on your phone when you have the cloud above you?

 

[KevinK]

Do smartphones really need the latest and greatest processor technology right out of the gate? If you have bloated software maybe.

You might be right in the long term. But right now LLMs and model environments continue to grow, just like smartphone software. Moore Threads can certainly build something that works for training and inference of current LLMs in 7nm or above, especially using chiplets. But the economics and power usage (TCOE) will strongly favor the highest on-chip and on-substrate densities, as well as placing the right amount of distributed memory close to the data pipelines. I think the Cerebras vs. Blackwell power comparison I posted earlier highlights the benefits of on-chip density and scale, as long as one can put and feed enough memory on the chip / wafer.

Thought this was a really interesting article on what's needed to tackle inference for very large models (1 trillion parameters) - lots of different forms of parallelism yield different tradeoffs of latency/interactivity vs throughput.

Demystifying AI Inference Deployments for Trillion Parameter Large Language Models | NVIDIA Technical Blog

AI is transforming every industry, addressing grand human scientific challenges such as precision drug discovery and the development of autonomous vehicles, as well as solving commercial problems such…

developer.nvidia.com