[Alpha Biz= Kim Jisun] Korean AI semiconductor fabless company FuriosaAI has begun mass production of its second-generation chip, Renegade (RNGD), marking a major step toward full-scale entry into the enterprise AI market.

FuriosaAI announced on the 28th that it has received the first mass-production batch of 4,000 RNGD units manufactured by its foundry partner TSMC. The chips will be shipped after completing card assembly through ASUS.

RNGD is offered in two configurations. The RNGD PCIe Card is designed as a low-power AI accelerator with a thermal design power (TDP) of 180 watts, allowing it to be installed directly into existing servers. The NXT RNGD Server is a 4U rack-mounted system equipped with eight RNGD cards, consuming only 3 kilowatts of total system power. Up to five servers can be installed in a standard rack, delivering a maximum of 20 PFLOPS (INT8) of AI inference performance per rack.

RNGD was first unveiled at Hot Chips 2024, a leading global semiconductor conference hosted by Stanford University in the second half of last year. The latest milestone represents one of the few cases globally in which a high-performance NPU (AI-dedicated semiconductor) equipped with HBM has progressed beyond development to full-scale mass production.

FuriosaAI has spent the past year stabilizing its hardware and enhancing its software stack. These efforts have led to tangible commercial progress, including confirmed deployment with LG’s EXAONE model and live demonstrations of OpenAI’s GPT-OSS models. The company has also established a stable mass-production and supply chain through close collaboration with global partners such as TSMC and SK hynix.

Built on an AI inference-optimized TCP architecture, RNGD delivers 2.5 times higher compute density per rack than GPU-based systems under standard operating conditions. This enables significantly greater AI inference workloads to be processed within the same space and power constraints.

 

 

 

[ⓒ 알파경제. 무단전재-재배포 금지]