FuriosaAI Targets AI Inference Market with Power-Efficient RNGD Chip

As AI model performance competition accelerates, the importance of efficient AI semiconductors and power infrastructure for real-world data centers is increasing. FuriosaAI has identified 'inference' as a key market in the large language model (LLM) era and is entering the AI infrastructure market with high-performance, low-power NPUs. At the GWDC 2026 Korea event held at Seoul's aT Center on March 30, Managing Director Jeong Young-beom introduced FuriosaAI's development of its self-developed AI semiconductor 'RNGD (Renegade)' and its performance and market strategy under the theme 'From Chip to Market.'
Jeong explained that since its founding in 2017, FuriosaAI has developed two types of AI semiconductors. The first generation focused on efficiently driving vision AI applications, as the AI market at that time centered around image and video processing. The goal was to create NPUs more efficient than NVIDIA GPUs for such workloads. The shift came with the emergence of GPT-3, which led FuriosaAI to predict LLMs would dominate the future AI market. They began developing a second-generation AI semiconductor optimized for LLMs.
RNGD incorporates high-bandwidth memory HBM3, designed based on the prediction that large language models will become dominant. RNGD was completed in 2024 and entered mass production this year, starting sales in January. The chip is manufactured using TSMC's 5nm process. FuriosaAI focuses on the AI 'inference' market, projecting that by 2030, over 100 gigawatts (GW) of data centers will be built globally, with approximately 70% dedicated to inference use.
Jeong highlighted RNGD's performance metrics, noting it supports 512 teraflops (TFLOPS) of computing power, capable of processing about 512 trillion operations per second. However, FuriosaAI emphasizes that competitiveness goes beyond raw computational speed; real-world data center efficiency depends on power consumption, memory, and application efficiency. With AI data center expansion making power supply a key constraint, 'performance per watt' has become a critical benchmark for AI semiconductors.
Jeong stated RNGD demonstrates higher power efficiency compared to NVIDIA GPUs, having proven its superior power efficiency (power efficiency). FuriosaAI also built the 'Furiosa LLM' software ecosystem to ease adoption, allowing users to run AI models on RNGD with minimal code changes from existing GPU environments. This strategy aims to reduce the burden of redeveloping entire AI services when adopting new NPUs.
Jeong emphasized that the AI semiconductor market is evolving from a race for faster chips to one focused on how efficiently AI services can be operated in real data centers. As LLM adoption grows, demand for computing resources and power will increase alongside, making comprehensive solutions combining performance, power efficiency, memory architecture, and software compatibility essential.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: [GWDC 2026] 퓨리오사AI "AI 경쟁, 칩 넘어 전력 효율로…RNGD로 추론 시장 공략"