FuriosaAI

According to IT Home, FuriosaAI, together with Broadcom, unveiled details of its third-generation AI inference accelerator, featuring 2nm compute dies, a separate I/O die, 12 stacks of 48GB HBM4 memory, PCIe Gen7, and 32 PFLOPS of performance. The chip is said to have a physical size eight times that of the second generation, with 576GB of HBM memory and 48TB/s bandwidth, representing 12x and 32x increases respectively compared to RGND."