FuriosaAI has revealed technical details of its third-generation AI inference accelerator, co-developed with Broadcom, featuring two 2nm compute dies, one I/O die, 12 48GB HBM4(E) memory stacks, PCIe Gen7 support, 32 PFLOPS compute, 576GB HBM capacity and 48TB/s bandwidth — representing 32x, 12x and 32x improvements over the second-gen RGND respectively.