Korean AI semiconductor startup FuriosaAI unveiled its third-generation AI inference accelerator developed with Broadcom. The chip features two 2nm computing dies, an independent I/O die, and 12 HBM4(E) memory stacks totaling 576GB. Utilizing SUE full interconnection topology and supporting PCIe Gen7, the new accelerator boasts 32 PFLOPS of computing power, a 12-fold increase in memory capacity, and 48TB/s bandwidth, representing a 32x performance leap over its second-generation predecessor.