ByteDance Denies Collaborating with Broadcom to Develop Dedicated Chips

On June 24th, it was reported that ByteDance is collaborating with Broadcom from the United States to develop advanced AI processors in order to reduce procurement costs and ensure stable supply of high-end AI chips. In response to this news, ByteDance stated that the information is not true.

According to media reports, the processor developed by ByteDance in collaboration with Broadcom is a 5nm ASIC (Application-Specific Integrated Circuit). ASIC is a type of specialized chip designed for specific application scenarios, achieving maximum performance and energy efficiency through algorithm hardening. It has strong average performance, low power consumption, and high cost-effectiveness, but requires large initial investment and long research and development time, and lacks versatility. It is expected that this processor will comply with the new US export control regulations, with part of the manufacturing work outsourced to TSMC.

Sources said that although the chip design work is currently progressing smoothly, it has not yet entered the ‘tape-out’ (trial production) stage, and it is expected that TSMC will not produce this chip this year. As of press time, TSMC has not yet responded. Previously, ByteDance had purchased Broadcom’s Tomahawk 5nm chip and Bailly switch for AI computing clusters.

ByteDance has been continuously investing in the chip field in recent years. Tan Dai, President of ByteDance’s subsidiary Volcano Engine, stated in a previous media interview that computing power is a heterogeneous computing system in the current industry situation. Each team has different chip choices, and different chips excel in different areas. Some have large memory, some have strong computing power, and some have good compatibility. For platform providers, scale advantages always exist. Regardless of whether more underlying suppliers will appear in the future, companies will consider cooperating with leading platforms.

SEE ALSO: OPPO Multiple Models Access ByteDance DouBao Large Model