FuriosaAI reveals third-gen AI inference accelerator specs with 32 PFLOPS

FuriosaAI has disclosed technical details of its third-generation AI inference accelerator. Developed in collaboration with Broadcom, the chip features two reticle-limit-class 2nm compute dies, one dedicated I/O die, and 12 HBM4(E) memory stacks of 48GB each, with SUE full-connection topology in the rack and PCIe Gen7 support.

According to FuriosaAI, the new chip is approximately eight times the physical size of its second-generation RGND chip, delivers 32 PFLOPS of compute (32 times that of RGND), 576GB of HBM capacity (12 times), and 48TB/s of HBM bandwidth (32 times).