FuriosaAI, a Seoul-based AI chip startup founded in 2017, is making waves in the AI hardware market with its innovative accelerator chips aimed at hyperscale data centers and enterprises. In August 2024, the startup unveiled its advanced AI chip, "RNGD." The chip is positioned to challenge industry leaders such as Nvidia with a 40% power efficiency advantage over Nvidia's H100 GPUs and competitive performance metrics. Furthermore, it recently gained attention for its testing and sampling progress with global enterprises like LG AI Research and Aramco.
Speaking to DIGITIMES Asia, FuriosaAI CEO June Paik shared insights into the company's journey, the capabilities of its AI accelerator, its recent collaboration with LG, its outlook for the AI landscape, and international collaborations.
Strategic foundations drive early success
At the company's inception, Paik and the co-founders had a long-term vision, targeting AI inference infrastructure for data centers. FuriosaAI initially focused on foundational chip design, delaying silicon production until it secured sufficient funding. "We treated design like writing the screenplay before making the movie—minimizing costs while perfecting the groundwork," Paik said.
The approach bore fruit with their first-generation chip, released in 2021. Despite a modest budget of US$6 million, it delivered a robust performance, achieving a 113% improvement in MLPerf benchmarks within weeks of its release. This success helped FuriosaAI secure a US$60 million Series B funding round, enabling the ambitious development of RNGD.
Betting on large language models, RNGD, developed in collaboration with Taiwan's Global Unichip Corporation (GUC) and TSMC, leverages cutting-edge technology, including HBM3 memory and a 5nm process node. FuriosaAI strategically designed the chip to excel in large-scale model inference, anticipating the rise of transformer models like GPT-3 and beyond.
"Our decision to bet on high-performance AI models in 2021 has paid off," Paik said. The company's early adoption of advanced technologies positions RNGD as a flexible and efficient solution for the growing demands of AI applications.
Innovation drives market differentiation
FuriosaAI's success lies in its holistic approach to chip design. Paik highlighted the startup's ability to balance performance, energy efficiency, and flexibility—key challenges in AI computing. The company's proprietary Tensor Construction Processor architecture underpins its innovative edge.
"We've worked for seven years to develop a scalable, sophisticated design," Paik said. "This fundamental innovation is what sets us apart from competitors like Nvidia and Intel."
As competition intensifies in the AI hardware market, Paik emphasizes the startup's focus on efficiency and flexibility positions it as a formidable challenger in the field.
Strategic partnership with LG AI Research
In August 2024, LG AI Research began integrating RNGD into its AI infrastructure, aiming to reduce dependency on Nvidia chips while advancing South Korea's AI semiconductor ecosystem. When asked about this pivotal collaboration, Paik highlighted the synergy between LG AI Research's goals and FuriosaAI's offerings.
"LG AI Research's mission is twofold: conducting advanced AI research and embedding AI into LG's diverse business operations," Paik explained. From manufacturing to product development across LG's numerous sectors, efficient AI solutions are imperative. Cost efficiency, in particular, plays a critical role in AI implementation in production environments, where high costs could hinder broader adoption.
RNGD offers a rare alternative to GPU-based solutions, capable of handling cutting-edge AI models with high efficiency. This unique value proposition attracted LG AI Research, positioning FuriosaAI as a key enabler in the transition toward more cost-effective and domestically sourced AI solutions.
Innovation reshapes industry landscape
The AI chip industry is fiercely competitive, dominated by established players like Nvidia. Yet FuriosaAI sees this as an opportunity to revolutionize the field with groundbreaking innovation. Paik likened the industry's evolution to the transition from gasoline-powered cars to electric vehicles.
"The shift to electric vehicles required not just better cars but also the infrastructure—charging stations, longer battery life, and more. Similarly, moving from GPU-based solutions to AI-specific chips requires strong software ecosystems to ease the transition for developers," he explained.
FuriosaAI's approach centers on building a robust software stack to support its hardware, enabling developers and users to transition seamlessly to its chips. RNGD embodies this philosophy, designed to prove that AI-specific chips can deliver superior efficiency and cost-effectiveness.
"Our goal is to be the first company to demonstrate that this transition is not only feasible but also advantageous," Paik asserted. He compared FuriosaAI's ambition to iconic electric vehicle pioneers who transformed the automotive industry by delivering reliable and sustainable alternatives.
Regarding the scale of the market, Paik underscored the massive potential of the AI chip market, which is expanding at an unprecedented pace. "The market is so vast that it can accommodate multiple players alongside dominant ones like Nvidia," he explained.
Global expansion strengthens Taiwan partnerships
Internationally, FuriosaAI is actively forging partnerships worldwide to expand its influence across key markets in the US, Asia, and beyond. Paik highlighted a Memorandum of Understanding signed with Saudi Arabia's Aramco in September, marking a strategic collaboration that underscores the company's commitment to global expansion. While details remain under wraps, FuriosaAI is engaging with potential customers across multiple regions, including the US, Japan, and India, reflecting its growing presence in the AI semiconductor space.
"We are excited about the current progress with our early customers. Our customer sampling pipeline is full and we're looking forward to working with global enterprises to deploy and integrate RNGD into their inference infrastructure. We expect to see first customer orders starting from the first quarter of 2025," Paik added.
Paik also emphasized the importance of Taiwan in FuriosaAI's ecosystem. As a hub for cutting-edge manufacturing, Taiwan plays a central role in the production of FuriosaAI's chips, including RNGD. Collaborations with Taiwan-based industry leaders have been pivotal in ensuring high-quality production and innovative server solutions. FuriosaAI plans to deepen its ties by potentially establishing an office in Taiwan to strengthen collaboration with the local ecosystem.
Credit: FuriosaAI
