FuriosaAI is jointly developing a next-generation AI inference platform with Broadcom. The two companies will combine FuriosaAI's AI semiconductor architecture with Broadcom's networking technology to develop a third-generation accelerator for hyperscale AI infrastructure.

Furiosa AI announced on the 28th that it has signed a strategic partnership with Broadcom and is pursuing the joint development of next-generation AI accelerators.

The two companies plan to advance FuriosaAI's Tensor Compound Processor (TCP) architecture into a multi-die based chiplet system and apply Broadcom's AI networking and high-bandwidth Ethernet switch technologies.

This development aims to build an inference infrastructure platform that integrates AI computing, networking, and software. It is a structure that considers not only computational processing but also the efficiency of data movement between servers and racks in a large-scale AI inference cluster.

The underlying product is FuriosaAI's second-generation AI accelerator, RNGD.

RNGD is a 180W PCIe AI accelerator based on TSMC's 5nm process and SK Hynix's HBM3, designed for large-scale language models and agentic AI inference workloads.

Verification work was also conducted in customer environments, including Samsung SDS and LG AI Research Institute.

The third-generation AI accelerator being developed by the two companies uses a 2-nanometer process-based compute die and HBM4 and HBM4E memory.

It is being developed to integrate multiple silicon dies into a single chip using Broadcom's advanced packaging technology and to support high-bandwidth rack-level networking.

Sampling of the next-generation AI accelerator is scheduled for the first half of 2028. Both companies are targeting the hyperscale inference infrastructure market to support the proliferation of ultra-large frontier models and agentic AI services.

Charlie Kawas, President of Broadcom’s Semiconductor Solutions Group, emphasized data reuse and communication efficiency between servers and racks, stating, “AI inference performance is no longer determined solely by simple computational performance.”

Baek Jun-ho, CEO of Furiosa AI, said, “We will provide a hyperscale AI inference platform by combining Broadcom’s infrastructure capabilities with Furiosa AI’s TCP architecture and software stack.”