[DigitalToday reporter Seok Dae-geon] AI semiconductor company FuriosaAI has entered mass production of its second-generation AI chip RNGD (Renegade). FuriosaAI said on Jan. 28 it received 4,000 RNGD chips from TSMC and will begin shipping them in card form from this week. Based on that, it aims to mass-produce 20,000 units this year and is moving to target the enterprise market.

RNGD was unveiled in the second half of last year at Stanford University's academic event Hot Chips 2024. It is a case in which an AI-dedicated semiconductor equipped with high bandwidth memory (HBM) has reached the mass-production stage. Chips completed by TSMC will be shipped after ASUS's card manufacturing process.

The RNGD PCIe card has a 180W low-power design and can be installed directly in existing servers. The NXT RNGD server is a 4U rack-mount server equipped with 8 RNGD cards, with total system power consumption of 3 kW. Up to 5 units can be installed in a standard rack environment, providing up to 20 PFLOPS (INT8) AI inference performance per rack.

According to Furiosa, one domestic conglomerate affiliate placed an order for an RNGD purchase in early January. FuriosaAI confirmed the adoption of LG EXAONE in the second half of last year. It also held a public demonstration of the OpenAI GPT-OSS model. It built a mass-production system in cooperation with TSMC, SK Hynix and others.

RNGD is based on a TCP architecture optimized for AI inference. In a standard environment, it provides 2.5 times higher compute density per rack than GPUs. It can handle more AI inference tasks under the same space and power conditions. FuriosaAI CEO June Paik (백준호) said, "Mass production of RNGD is a step forward toward becoming one of the top three global AI players and a top-two semiconductor player," and added, "We will accelerate efforts to expand sales in global markets."