Meta in talks to acquire Korean AI chip startup FuriosaAI
Key Points
- 1Meta is reportedly in discussions to acquire FuriosaAI, a South Korean AI semiconductor startup, with a deal potentially closing this month as reported by Forbes.
- 2FuriosaAI specializes in AI NPUs for datacenters, with their latest chip, Renegade (RNGD), offering three times the AI computation per watt compared to Nvidia H100 and being optimized for Meta's Llama 2 and 3 models.
- 3This acquisition would significantly bolster Meta's AI semiconductor competitiveness, leveraging FuriosaAI's specialized architecture for large language models.
Forbes reported on February 11, citing anonymous sources, that Meta is in discussions to acquire FuriosaAI, a South Korean AI semiconductor startup. The article, published on February 13, indicates that negotiations are ongoing, with a deal potentially closing as early as the current month. However, the outcome remains uncertain as multiple companies have expressed interest in acquiring FuriosaAI. Both FuriosaAI and Meta have refrained from immediate comment, with FuriosaAI stating they cannot respond to speculative reports.
Established in 2017 by CEO Baek Jun-ho, who holds a Master's degree from Georgia Tech and previously worked as a GPU/CPU development engineer at Samsung Electronics and AMD, FuriosaAI is a fabless AI semiconductor company. It specializes in developing Neural Processing Units (NPUs) optimized for AI computation in data centers and autonomous driving markets. The company has attracted significant investment due to its advanced technology, accumulating 170 billion KRW in total funding from major investors including Naver, DSC Investment, and Crit Ventures. FuriosaAI has also been selected as a project partner by the Korean Ministry of Science and ICT for its AI semiconductor initiatives.
FuriosaAI unveiled its first AI semiconductor, 'Warboy,' in 2021, followed by its next-generation chip, 'Renegade' (RNGD), in August of the previous year. According to FuriosaAI, RNGD offers significant performance advantages, capable of processing three times more AI computation with the same power consumption compared to NVIDIA's H100. A key feature of RNGD is its optimization for large language models (LLMs) such as LLaMA 2 and LLaMA 3. The company emphasizes that the process of optimizing and executing AI models on GPUs is often complex, time-consuming, and demands high expertise, particularly with the emergence of new chip architectures that necessitate kernel tuning and intricate adjustments during deployment. RNGD addresses this challenge with its tensor-based hardware architecture, designed to automatically deploy and optimize new AI models, even those with unique structures. Furthermore, unlike conventional high-performance GPUs, RNGD can be deployed in various data centers without requiring complex liquid cooling systems.
Forbes analysts suggest that the potential acquisition would significantly enhance Meta's AI semiconductor competitiveness, given RNGD's optimization for Meta's open-source LLaMA 2 and LLaMA 3 models.