South Korea's FuriosaAI Reveals Third-Gen AI Inference Chip with 32 PFLOPS

South Korean AI chip startup FuriosaAI has shared details about its third-generation AI inference accelerator. The chip is co-developed with Broadcom and features two 2nm compute dies, one separate I/O die, and 12 stacks of 48GB HBM4 memory. It supports PCIe Gen7 and uses a SUE fully-connected topology within the rack.

FuriosaAI says the chip is about 8 times the physical size of its second-gen chip (RGND), with computing power reaching 32 PFLOPS, 32 times that of RGND.

HBM memory capacity hits 576GB, 12 times that of RGND, while memory bandwidth reaches 48TB/s, also 32 times that of RGND. The chip targets large-scale AI inference with major performance gains.