ByteDance Develops AI GPU to Reduce Dependence on Nvidia

ByteDance is developing two GPUs dedicated to the field of artificial intelligence, and plans to achieve mass production by 2026. And these two AI GPUs will be manufactured by TSMC. It will help ByteDance reduce dependence on NVIDIA and comply with the export control regulations of the United States.

Currently, these two GPUs are in the design phase. ANd one is specifically designed for AI training and the other for AI inference. They will be produced using TSMC’s advanced process technology N4 or N5, similar to Nvidia’s Blackwell series process technology. ByteDance may realize the large-scale production and deployment of these GPUs in 2026.

This year, ByteDance has invested more than $2 billion to purchase more than 200000 Nvidia H20 GPUs. And the unit price is about $10000. But many of them have not yet been delivered. Due to the shortage and high price of NVIDIA GPUs, ByteDance decided to develop its own AI hardware.

Challenges

Nvidia has specially designed products such as DGX H20 for the Chinese market in response to the export control policies implemented by the United States last year. The performance of HGX H20 has been reduced compared to Nvidia’s high-end H100. But it still comes with 96GB of HBM3 memory, a maximum memory bandwidth of 4TB/s, and 8-way GPU interconnect capability. That makes it still favored by enterprise customers in practical applications. And the new GPU with ByteDance may be restricted by US export control and cannot surpass HGX H20 in performance. But its cost will drop greatly.

ByteDance faces a biggest challenge in self-development of AI GPU. That is the company currently relies on Nvidia’s CUDA and corresponding software stack for AI training and reasoning. If switching to their own GPU, ByteDance will require the development of new software platforms. Moreover, it will ensure the compatibility of software and hardware.