ByteDance is developing its own GPU for AI, likely to be manufactured by TSMC to reduce its dependence on Nvidia and comply with US export regulations.
According to sources from Tomshardware, ByteDance is in the process of designing its own GPU for artificial intelligence (AI), with the aim of reducing its dependence on Nvidia, the leading GPU supplier today. This move is said to be to cope with the shortage of GPUs as well as increasingly strict export control regulations in the US.
ByteDance is reportedly designing two types of GPUs, one for AI training and one for AI inference. Broadcom, which previously designed AI chips for Google, is likely to continue designing chips for ByteDance. The GPUs will be manufactured on TSMC's N4/N5 process, similar to Nvidia's Blackwell GPU line for AI and HPC. ByteDance's AI GPU products are expected to be deployed around 2026.
While ByteDance’s GPUs are expected to be cheaper than Nvidia’s, their performance competitiveness may be limited by US export controls. TSMC may not be allowed to produce chips that outperform Nvidia’s current HGX H20 GPUs, making it difficult for ByteDance to outperform its rivals in processing speed.
This year, ByteDance invested more than $2 billion to purchase more than 200,000 Nvidia H20 GPUs, each costing about $10,000. This large investment shows the importance of AI to ByteDance's overall strategy. However, the shortage of Nvidia GPUs and high prices have prompted ByteDance to develop its own AI hardware solution.
Being self-sufficient in AI GPU production is expected to help ByteDance control costs more effectively, especially when Nvidia GPUs, even though they have been tweaked for the Chinese market, are still expensive.
Nvidia CEO Jensen Huang. The company expects to ship more than 1 million of its new H20 chips to China this year, priced at $12,000 to $13,000 each. Source: Nvidia. ByteDance’s software ecosystem challenge
However, ByteDance also faces the challenge of developing a software ecosystem for its new AI GPUs. Currently, the company relies on CUDA and Nvidia's support tools for AI training and inference.
Switching to its own GPU means ByteDance needs to build a complete software platform and ensure compatibility with new hardware, a difficult problem not only for ByteDance but also for many other Chinese companies trying to develop their own AI GPUs. Relying on Nvidia solutions for some tasks is also a common situation for many domestic companies.
Nvidia is expected to ship more than 1 million HGX H20 GPUs to the Chinese market this year, nearly double Huawei's sales of its Ascend 910B AI GPU series, which is expected to reach 550,000 units in 2024. Revenue from HGX H20 GPUs could bring Nvidia more than $12 billion, surpassing its total revenue from the Chinese market in the previous year.