FuriosaAI and Broadcom announce strategic partnership to develop third-generation AI accelerator for agentic inference at scale.

FuriosaAI and Broadcom are combining FuriosaAI's Tensor Contraction Processor (TCP) architecture and software stack with Broadcom's scale-up AI networking solutions to build a rack-scale inference platform for serving frontier agentic systems.

FuriosaAI's second-generation chip, RNGD, is a 180W, PCIe-based accelerator fabricated at TSMC's 5nm process and currently in mass production, validated by Samsung SDS and LG AI Research. The third-generation chip will feature a 2nm compute die, a dedicated IO die for scale-up networking, and HBM4/4E memory, integrated using Broadcom's advanced packaging capabilities.

The partnership aims to deliver industry-leading performance per watt for large, complex frontier AI models and agentic workloads (AI systems that make decisions and take actions autonomously) by addressing key bottlenecks of large-scale agentic AI through low-latency, high-bandwidth all-to-all interconnect across hundreds of chips at rack scale.