Korea AI semiconductor startup FuriosaAI is joining hands with global semiconductor corporations Broadcom to develop a next-generation AI inference platform.
FuriosaAI said on the 28th that it signed a strategic partnership with Broadcom to jointly develop a next-generation AI accelerator.
The two sides plan to advance FuriosaAI's proprietary chip architecture, the TCP (Tensor Contraction Processor), into a multi-die chiplet system to build a next-generation platform that meets the demand for large-scale token processing in global hyperscale AI environments.
The platform will be developed based on FuriosaAI's second-generation AI accelerator, "RNGD." RNGD is a 180W PCIe-based AI accelerator that applies TSMC's 5-nanometer process and SK hynix HBM3, and it is said to be optimized for large language models (LLMs) and agentic AI workloads.
FuriosaAI said it has completed RNGD validation in customer environments in and outside Korea, including Samsung SDS and LG AI Research Institute, and that its related partner ecosystem is expanding rapidly.
The third-generation AI accelerator to be co-developed by the two sides will feature compute dies built on a 2-nanometer process and HBM4 and HBM4E memory. In addition, Broadcom's advanced packaging technology will be applied to integrate multiple silicon dies into a single high-performance chip.
The two sides are aiming to begin sampling the next-generation AI accelerator in the first half of 2028.