China, 4China Mobile Research Institute, Beijing, China, Xiamen Industrial Technology Research Institute, Fujian, China, 6Bytedance China, Beijing, China, 7Huawei Technologies, Shenzhen, China 1 5 Abstract We present the first 28nm hybrid compute-in-RRAM (CiR) accelerator for recommendation system (RecSys) based on the HYDAR framework: (1) DL ADCs enable early termination of non-Top-K calculations. (2) PPSP dataflow boosts throughput for irregular workloads. (3) A coarse-to-fine architecture preserves system recall accuracy. The 36M-RRAM CiR chip achieves 390K QPS with a SOTA 1574K QPS/W. The chips scale out to a 576M system for practical RecSys task, improving by 66x in QPS and 181x in QPS/W. Recommendation systems (RecSys) play a crucial role in connecting users with massive
Jiaming Li1, Peng Yao1, Xueqi Li1, Zhenqi Hao1, Dabin Wu1, Zhouzheng Li1, Haishu Xianyu2, Lin Li3, Shujuan You4, Taiwei Chiu5, Maojia Sheng6, Wei Yang7,
Qingtian Zhang1, Jianshi Tang1, He Qian1, Bin Gao1, Huaqiang Wu1 Tsinghua University, Beijing, China, 2Beijing Elemem Technology, Beijing, China, 3Migu Culture Technology, Beijing,