德拉姆
歪斜
校准
计算机科学
模具(集成电路)
窗口(计算)
基础(拓扑)
可靠性(半导体)
电子工程
工程类
序列化
芯(光纤)
计算机硬件
立方体(代数)
作者
Sunghwan Joo,Jinyeon Kim,Yongsun Lee,Ji-Young Kim,Youngsik Lee,Yongmin Kim,Chisung Oh,Kyu-Ha Shim,Haesuk Lee,Young-Yong Byun,ChangHyun Bae,Joohwan Kim,Je-Min Ryu,Shin-haeng Kang,Jaehoon Lee,Young-Uk Chang,JaeKyung Lee,Jongtae Kim Hwang,Daehwan Seo,K. Na
标识
DOI:10.1109/isscc49663.2026.11409139
摘要
LLM-based generative AI requires high-capacity and high-bandwidth memory. We propose an HBM4 with 2048 I/Os, 3.3TB/s bandwidth, and 36GB cube density. Core die uses the 6th gen. DRAM and the base die uses a 4nm Fin-FET process. TSV count and data window are doubled for high-frequency margin. tCCDR is improved via per-channel TSV-RDQS auto calibration, enhancing rank-to-rank access. PMBIST enables at-speed testing in CoW and SiP. The on-die WDQS skew cancellation ensures robust timing, enabling reliable highspeed operation and improved yield.
科研通智能强力驱动
Strongly Powered by AbleSci AI