South Korea's FuriosaAI Unveils Second-Gen AI Inference Chip, Outperforming Nvidia RTX PRO6000 in Energy Efficiency by Multiple Times

South Korean AI chip company FuriosaAI recently held its first RENEGADE summit, announcing multiple progress updates on its high-efficiency data center AI inference chip product line.

The second-gen RNGD chip, when running the Qwen3-32B (FP8) model under the same power consumption, can support 2.2 to 7.4 times more inference service users than Nvidia's RTX PRO6000, with a total cost of ownership (TCO) only 40% of the latter. 4,000 RNGD chips were mass-produced in January this year, and the company plans to reach a total annual capacity of 20,000 units. In the short term, it will upgrade memory from 48GB HBM3 to 72GB HBM3E to enhance performance.

Looking ahead, FuriosaAI intends to launch a third-gen chip in 2028 to further improve performance; it is also planning a lightweight chip RNGD S for PCs and workstations, expected to be released from late 2026 to early 2027.