SqueezeBits-Naver-KAIST Joint Research on Intel AI Accelerators Accepted for ISCA 2025

Domestic AI startup SqueezeBits, together with Naver Cloud and KAIST Professor Yoo Minsu's lab, has proudly been listed at ISCA 2025, the world's most prestigious computer architecture conference, with a paper analyzing the performance of Intel's AI accelerator 'Gaudi-2'. Celebrating its 52nd anniversary this year, ISCA is renowned as a highly selective conference in the field of computer architecture, with an acceptance rate of less than 20%. The research team meticulously analyzed Gaudi-2's performance and power efficiency in various AI environments and conducted a comparative evaluation with NVIDIA GPUs (A100). Particularly in large language model (LLM)-based service environments, Gaudi-2 attracted attention by demonstrating an average of 50% better power efficiency compared to NVIDIA's A100. On the other hand, in recommendation system (RecSys) environments where access to multiple user data and simple computations are repeated, it recorded somewhat disappointing results in terms of performance and power efficiency compared to the A100. This research was conducted as a collaborative project of the 'AI Joint Research Center (NICL)' established by Naver Cloud and Intel. It is a result of strategic cooperation where Naver Cloud supports startups and academia in utilizing Intel Gaudi-based AI infrastructure. Particularly, it is significant because it further enhanced the trustworthiness of the technology through independent verification by a third party (startups and academia), rather than Intel's self-assessment as the manufacturer. SqueezeBits, which attracted investment from Naver D2SF in 2022, is leading the development of Intel Gaudi-2-based generative AI inference technology through close cooperation with Naver. Kim Hyung-jun, CEO of SqueezeBits, stated, "We are very pleased that our technology has been officially recognized by a global academic conference," and expressed his ambition, "Moving forward, we will continue to contribute to enhancing the competitiveness and diversity of the AI semiconductor ecosystem together with our collaborating institutions." SqueezeBits, a startup specializing in AI model lightweighting and optimization, currently leads the market by providing the AI model lightweighting optimization toolkit 'OwLite' and the large language model (LLM) serving and deployment optimization tool 'Fits on Chips' as SaaS. This year, it plans to expand its collaborations with leading AI semiconductor companies, including Rebellions.