FuriosaAI unveils new AI inference accelerator with 32x computing power and memory bandwidth

On September 21, 2026, South Korean AI semiconductor startup FuriosaAI announced the technical specifications of its third-generation AI inference accelerator, developed in collaboration with Broadcom. The chip features two 'mask limit-level' 2nm compute dies, one independent I/O die, and 12 stacks of 48GB HBM4 (E) memory. It employs a SUE fully connected topology and supports PCIe Gen7. FuriosaAI stated that the chip's physical size is approximately eight times larger than its second-generation model, RGND. The new accelerator delivers 32 PFLOPS of computing power, 32 times that of RGND, with 576GB of HBM memory, 12 times the capacity of RGND, and a memory bandwidth of 48TB/s, 32 times that of RGND. The company emphasized the significant improvements in performance and efficiency, positioning the new accelerator as a major advancement in AI inference technology. FuriosaAI's collaboration with Broadcom highlights the growing trend of partnerships in the AI semiconductor industry, aiming to push the boundaries of computational capabilities for machine learning applications.