en.Wedoany.com Reported - FuriosaAI announced a strategic partnership with Broadcom to jointly develop a third-generation AI accelerator platform based on a multi-chip chiplet architecture, specifically optimized for large-scale inference workloads. The collaboration combines FuriosaAI's Tensor Contraction Processor architecture with Broadcom's AI networking technology, high-bandwidth Ethernet switching, PCIe connectivity, and advanced packaging capabilities to build a rack-scale inference platform for hyperscale AI deployments.
The new platform surpasses FuriosaAI's existing RNGD inference accelerator, which is currently in mass production and manufactured using TSMC's 5nm process. RNGD is a 180W PCIe accelerator designed for large language model inference and agentic AI workloads operating in standard air-cooled data center environments, with production deployment validated by organizations such as Samsung SDS and LG AI Research. The third-generation architecture will advance to a chiplet-based system, featuring 2nm compute dies and dual-layer HBM4/4E memory.
Broadcom will provide Ethernet scaling technology, PCIe interconnect IP, packaging integration, and AI network switching infrastructure, supporting the scaling of inference clusters to thousands of nodes. FuriosaAI stated that the architecture emphasizes high-bandwidth data transfer and communication efficiency, rather than traditional GPU thread management approaches. The software stack includes a compiler-driven software development kit that maps high-level PyTorch workloads directly onto the chip, and features a virtual ISA abstraction layer that exposes underlying hardware control to developers, eliminating the need to manage traditional GPU kernel optimization workflows.
FuriosaAI is adopting a chiplet-based multi-chip architecture to design its third-generation accelerator, combining 2nm compute dies with dual-layer HBM4/4E memory. The RNGD inference accelerator remains in mass production on TSMC's 5nm process, consuming 180W in a standard PCIe server configuration. The architecture is geared towards large-scale inference clusters supporting high-throughput token generation, focusing on communication efficiency, memory bandwidth, and rack-level scalability. FuriosaAI's compiler-based software development kit aims to reduce reliance on manually tuned kernels.
Charlie Kawwas, President of Broadcom's Semiconductor Solutions Group, stated that inference performance is no longer determined solely by raw compute power, but increasingly depends on data reuse and communication efficiency across servers and racks. By combining Furiosa's Tensor Contraction Processor architecture with Broadcom's XPU technology, IP platform, Ethernet scaling, and network switches, the two companies are building a platform that addresses critical bottlenecks in large-scale agentic AI. FuriosaAI CEO June Paik stated that integrating Broadcom's infrastructure capabilities with its own Tensor Contraction Processor architecture allows them to move beyond the chip level and provide a comprehensive solution for the era of token factories. The second-generation chip, RNGD, has been mass-produced by TSMC and has validated its performance and efficiency, and the third-generation inference solution will deliver industry-leading performance-per-watt for the largest and most complex frontier AI models.
This collaboration highlights the shift from monolithic AI accelerators to chiplet-based inference architectures tightly coupled with high-bandwidth Ethernet networking and advanced packaging. As hyperscalers advance towards continuous token generation for agentic AI systems, network efficiency, memory bandwidth, and rack-level communication are increasingly defining deployment economics alongside raw compute performance. The partnership also reinforces Broadcom's role in AI infrastructure, where it has steadily expanded its AI platform strategy to encompass network fabrics, advanced packaging, interconnect IP, and scale-out infrastructure. FuriosaAI joins a challenger lineup focused on inference, companies seeking to differentiate from GPU-centric architectures through specialized chips optimized for inference efficiency and lower power consumption.
The collaboration between FuriosaAI and Broadcom follows a significant funding milestone for the company. In July 2025, FuriosaAI completed a $125 million Series C bridge funding round, bringing its total funding to approximately $246 million. The round included support from Korea Development Bank, Industrial Bank of Korea, Keistone Partners, PI Partners, and Kakao Investment, along with participation from numerous institutional investors. FuriosaAI stated that the funds will be used to scale production of its RNGD inference processor while accelerating the development of its next-generation architecture, enabling it to remain independent and expand globally in the increasingly competitive AI inference chip market.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
