Reducing reliance on NVIDIA, ByteDance reportedly partners with Taiwan Semiconductor to independently develop 5nm AI chips, set to mass produce by 2026

Author: Li Dan

Source: Hard AI

Competition in the field of artificial intelligence (AI) is fierce. The demand for NVIDIA AI chips that power large language models (LLMs) is soaring, while ByteDance may introduce its own chips to reduce reliance on NVIDIA products.

According to tech media The Information, citing sources familiar with the matter, ByteDance is collaborating with TSMC, a key partner in NVIDIA's advanced chip manufacturing, to jointly produce two self-developed AI chips based on 5-nanometer (nm) technology by 2026. The project is currently in the design stage, and ByteDance's plans for the future may still change.

The report mentions that the 5nm technology is only one generation behind TSMC's most advanced 3nm technology, which is the same generation used in NVIDIA's upcoming flagship AI chip - the Blackwell architecture chip. Developing their own chips may reduce ByteDance's reliance on expensive NVIDIA chips for developing and operating AI models.

Amid intense domestic competition and strengthened export restrictions by the United States, reducing chip costs has become crucial for Chinese tech companies like ByteDance that are introducing LLMs. The report notes that ByteDance has already ordered over 200,000 NVIDIA H20 chips worth over $2 billion and is still awaiting full delivery. ByteDance plans to have TSMC manufacture hundreds of thousands of self-developed training and inference chips, which will cost tens of billions of dollars less than purchasing from NVIDIA.

As to why ByteDance may significantly reduce costs, the report points out that the performance of H20 is only one-fourth that of NVIDIA's popular advanced chip H100. The current U.S. export restrictions target individual chip performance but do not restrict chip clusters in this way. Domestic manufacturers need to purchase more H20 chips to achieve the same cluster effect as a single H100 chip. ByteDance hopes that, under roughly the same cost, clusters driven by self-owned chips can achieve four times the computing performance of a single H100 chip.

Both ByteDance and NVIDIA have not responded to the above reports. Earlier reports indicated that major Chinese internet companies, including ByteDance, have made significant orders for NVIDIA chips this year while actively developing self-developed AI chips.

Last month, the Financial Times reported that Baidu, ByteDance, Tencent, and Alibaba collectively ordered $1 billion worth of A800 processors from NVIDIA, with these Chinese companies also purchasing $4 billion worth of GPUs, all to be delivered this year.

A ByteDance spokesperson later did not provide detailed comments on the above reports but stated, "Consumer internet companies and cloud suppliers invest billions of dollars annually in data center components, usually placing orders months in advance."

In June this year, Reuters learned that ByteDance is collaborating with Broadcom to develop advanced 5nm AI chips, which are ASIC chips manufactured by TSMC and comply with U.S. export restrictions. At the time, Reuters reported that Broadcom is at least ByteDance's commercial partner until 2022, openly stating that ByteDance purchases its 5nm high-performance switch chips and the Bailly switch for AI computing clusters. The chip development cooperation between ByteDance and Broadcom will help reduce procurement costs and ensure a stable supply of high-end chips This week, The Information also reported that Baidu's next-generation AI chip Kunlun 3 is currently in the final stage of design development and is about to enter production at Taiwan Semiconductor