South Korean startup FuriosaAI has unveiled the technical specifications for its third-generation AI inference accelerator. The new hardware was developed in collaboration with Broadcom and utilizes a multi-die architecture.
The accelerator features two 2nm compute dies, one dedicated I/O die, and twelve 48GB HBM4 (E) memory stacks. The system supports PCIe Gen7 and employs a SUE all-to-all topology within the rack.
Compared to the previous generation, known as RGND, the new chip offers significant performance gains. The physical size of the new processor is approximately eight times larger than its predecessor.
The device delivers 32 PFLOPS of computing power, which is 32 times higher than the RGND model. Additionally, the chip provides 576GB of HBM capacity and a memory bandwidth of 48TB/s, marking a 32-fold increase in bandwidth.
