[Digital Today reporter Daegeon Seok] FuriosaAI is moving to jointly develop a next-generation AI accelerator with Broadcom. FuriosaAI said on May 28 it signed a strategic partnership with global semiconductor company Broadcom.

The companies plan to advance FuriosaAI’s proprietary chip architecture, the Tensor Contraction Processor (TCP), into a multi-die chiplet system. They also plan to jointly develop a next-generation AI inference platform to meet token processing demand in global hyperscale AI environments.

The cooperation aims to build an infrastructure platform integrating AI computing, networking and software, beyond joint semiconductor development. It will combine FuriosaAI’s AI architecture technology with Broadcom’s AI networking and high-bandwidth ethernet switch technology to form an integrated platform supporting large-scale AI inference clusters.

The next-generation platform is based on FuriosaAI’s second-generation accelerator, RNGD (Renegade). RNGD is a 180W PCIe AI accelerator based on TSMC’s 5-nanometer process and SK Hynix’s HBM3, and handles large language model (LLM) and agentic AI workloads. It has completed verification in customer environments including Samsung SDS and the LG AI Research Institute, and adoption by customers at home and abroad and expansion of the partner ecosystem are under way.

The third-generation AI accelerator will apply a 2-nanometer process-based compute die and HBM4/4E memory. It will integrate multiple silicon dies into a single chip using Broadcom’s advanced packaging technology, and support rack-level high-bandwidth networking in large-scale AI clusters by combining Broadcom’s ethernet and high-speed switch technologies. The companies plan to begin sampling the third-generation accelerator in the first half of 2028.

Charlie Kawwas (찰리 카와스), president of Broadcom’s Semiconductor Solutions Group, said, "AI inference performance is no longer determined simply by compute performance alone." He said, "Data reuse and communication efficiency between servers and racks are now key competitive strengths." He added, "We will build a platform that resolves key bottlenecks in large-scale agentic AI environments by combining FuriosaAI’s TCP architecture with Broadcom’s XPU technology and IP platform and ethernet scale-up networking technology."

June Paik (백준호), CEO of FuriosaAI, said, "As Broadcom’s infrastructure capabilities combine with FuriosaAI’s TCP architecture and software stack, we have become able to provide a hyperscale AI inference platform for the era of the 'Token Factory.'" He added, "In next-generation products, we will deliver industry-leading performance per watt even in ultra-large AI models and hyperscale agentic AI environments."