1、 Information Classification:General 112G Signal Integrity Investigation for Immersion-Cooling AI Server Pengfei Cheng,Douyin Vision Co.,Ltd, Kai Wang,Intel corporation,K Ying Zhang,Douyin Vision Co.,Ltd, Yulong Wang,ByteDance, Chen Shen,ByteDance, Hui Zou,ByteDance, Huaguo(Andy)Yu,IEIT, Wenbin Tian,
2、Intel,Tian, Akinwale,Oluwafemi,Intel, Zhiming Li,Intel, Xu U Zhang,Intel,Zhang, Paul Xu,Intel, Xiaoguo Liang,Intel, Ken Zhang,Intel, Nishi Ahuja,Intel, Information Classification:General Abstract This paper explores the benefits of immersion-cooled AI data centers,such as achieving a target PUE of 1
3、.1,excellent WUE water usage effectiveness,supporting high power density deployments,and ensuring high GPU and CPU reliability.However,immersion cooling poses signal integrity(SI)design challenges for 112G and 224G PAM4 interconnects.The dielectric constant(Er)of immersion liquids,typically between
4、1.8 and 2.2,leads to changes in connector impedance(OSFP,QSFP,Backplane connector and card edge connector)by 10 to 20 ohms and shifts resonance to lower frequencies.This paper presents measurement data for 112G PAM4 in an immersion cooling environment by 1)measuring Tx compliance,eye diagram,and var
5、ies channels BER on an NIC card,and 2)evaluating BER using a Serdes evaluation kit to drive long reach(LR)channels in immersion.It also demonstrates the use of channel operation margin(COM)analysis methods to examine the effects of liquid dielectric constant(Er)in immersion cooling.Additionally,it s
6、hows how an impedance-tuned QSFP connector for immersion can improve BER and COM performance.Author(s)Biography Pengfei Cheng is a SI Engineer of Douyin Vision Co.,Ltd.since2021.He focuses on the high speed and power integrity in server over the last decade.Kai A Wang.Intel Oregon,focus on AI and GP