Various services using AI are becoming mainstream, and as the AI model size increases, the service range is also expanding. Accordingly, it requires more computing performance and more memory capacity. The energy efficiency of AI computer system is fairly poor comparing with that of human brain. As a countermeasure against it, in this talk, Processing in Memory will be presented as one of the solutions. PIM architecture basically enables higher performance and lower energy consumption when performing memory intensive workloads. The current trending transformer based generative deep learning model, such as GPT2/3 shows memory intensive characteristics. The data analytics pipeline that pre-process and supplies data to the AI model also has a memory intensive feature. So, it is expected that PIM technology can be applied to the overall AI service computing system. In this talk, we’d like to introduce not only SK hynix’s 1st PIM product, GDDR6-AiM, but also CXL memory card based PIM solution and storage level PIM solution.

- 학력· 2003 ~ 2006 성균관 대학교 전자전기공학과 박사· 1993 ~ 1995 연세대학교 전자공학과 석사· 1989 ~ 1993 연세대학교 전자공학과 학사 - 주요경력· 2021 ~ 현재 SK 하이닉스 Fellow, Memory Solution Product Design· 2017 ~ 2021 SK 하이닉스 Fellow, System Architecture· 2009 ~ 2016 삼성전자 수석연구원, SoC Architecture· 2001 ~ 2009 삼성전자 책임연구원, SoC Design· 1995 ~ 2001 삼성전자 선임연구원, SoC Development