Samsung Electronics and South Korea's renowned search engine firm, Naver, have unveiled the outcomes of their year-long partnership in AI semiconductors, showcasing exceptional power efficiency. When pitted against commercial products from Nvidia and Google in a comparative test driving the Meta AI model LLaMa at equivalent speeds, the South Korean AI semiconductor displayed a staggering eight-fold superiority in energy consumption performance over its Western counterparts.
Building on this remarkable success, both companies are focusing on customized integrated circuits (IC) tailored for Naver Cloud. The collaborative effort, initiated with a business agreement signed in December 2022, was strategically aimed at investing in local AI semiconductor research. The ultimate goal was to establish a competitive edge against formidable overseas players such as Nvidia and Google.
As reported by The Korea Economic Daily and Theelec, the groundbreaking results of this collaboration, specifically in the realm of Field Programmable Gate Array (FPGA) AI semiconductors, were jointly unveiled by Samsung and Naver at the "4th Top-Level Strategic Dialogue on AI Semiconductors" event. This gathering, organized by the South Korean Ministry of Science and ICT, served as the platform for the two companies to demonstrate the AI semiconductor's remarkable capability in driving Naver's large-scale AI model, HyperCLOVA X.
A spokesperson from Naver explained that, during computation processes, rival products typically consume 16 bits to express a numerical value. In contrast, Naver and Samsung's products efficiently utilize only 4 bits, with minimal impact on overall performance.
As semiconductor density increases, industry attention is focused on whether the same computational efficiency can be achieved with less power. Both companies plan to integrate the developed AI semiconductors into their own services and commercialize them in a B2B product format.
However, while the FPGA product serves as a prototype for technical verification of Samsung and Naver's AI solution, it cannot be directly applied to Naver Cloud products. Therefore, Naver is collaborating with Samsung to develop an application-specific integrated circuit (ASIC) for large language models (LLM). Although the specific development direction is yet to be determined, the collaboration is expected to use Samsung's advanced manufacturing processes and incorporate eight LPDDR5X to target the field of AI application.
Kim Yuwon, CEO of Naver Cloud, pointed out that Naver has been working on creating a new matrix calculator and computational models for AI solutions. While FPGA serves the purpose of technology verification, its application becomes challenging when delivering large-scale AI model services. Consequently, Naver foresees an extended partnership with Samsung to design ASIC, with substantial time expected for the mass production phase.