Keisuke Kishishita

Socionext, Japan

How we design AI accelerator SoC for Data Center Network

Abstract

In recent years, data centers have become an essential infrastructure for our information society. At the same time, electricity consumption in data centers continues to increase, and the resulting CO2 emissions are a major challenge on a global scale. The increase in power is due to the huge amount of computation required for training and inference in AI, which is rapidly becoming popular. With the advent of generative AI, the amount of computation required to train AI increased 7 million times in 9 years from 2015 and is expected to continue. The increase in computational complexity has spurred power consumption, and there is a need to develop specialized semiconductor devices to reduce power growth. Semiconductor design has been striving to achieve high energy efficiency in System on a Chip (SoC) by leveraging technologies such as high-density CPUs, advanced packaging, HBM memory, high-speed interfaces, and silicon photonics. On the other hand, SoCs in the fields of AI accelerators, multi-core CPUs, and 5G networks, which are designed using these technologies, often become large-scale SoCs at advanced technology nodes, making their design extremely challenging. For such SoCs, integrated technologies that fuse multiple cutting-edge technologies are required to meet the demanded performance. In my talk, I will introduce these examples.

Biography

Keisuke Kishishita received a bachelor’s degree in electronics and information engineering from Kyoto Institute of Technology in 1999 and a master’s degree in design engineering and management from Kyoto Institute of Technology in 2017. In 1999, he joined Panasonic Corporation and worked on semiconductor design, including the back-end design of SoCs and SRAMs for mobile phones and home electrical devices. Since 2015, he has been in charge of the back-end design of high-end SoCs used in 5G networks, data centers, autonomous driving, etc., and is currently in charge of the business strategy of SoCs for data center and networking field.

Download