FuriosaAI Partners with Broadcom to Build Next-Generation Inference Platform for the Agentic AI Era

As AI evolves to be agent-centric and rapidly shifts toward inference-centric infrastructure, next-generation engines with massive token processing and compute capabilities are required. FuriosaAI has formed a strategic partnership with Broadcom to evolve its TCP (Tensor Contraction Processor) architecture into a multi-die chiplet system, developing a 3rd-generation AI accelerator.

This collaboration focuses on combining Broadcom's XPU Technology, IP platform, and Ethernet scale-up and fabric switch technology to address bottlenecks in large-scale agentic AI.

FuriosaAI has already proven its technological capabilities with RNGD, its 2nd-generation datacenter inference chip currently in mass production at TSMC. It also offers an alternative to CUDA through the Furiosa SDK, leveraging a general-purpose compiler and Virtual ISA to enable developers to rapidly deploy new models and optimizations.

The 3rd-generation platform roadmap includes the following technical advances:

• Application of HBM4/4E and 2nm process technology
• Support for high-speed inter-chip networking and all-to-all topology
• Optimization of complex communication patterns such as MoE (Mixture-of-Experts) routing