2020.07装不下GPU Memory Wall / Training Memory Footprint
2021.01分层卸载CPU Offloading
2021.04无限内存Heterogeneous Memory Hierarchy
2022.05IO-AwareIO-Aware Exact Attention
2022.06内存池化CXL Tiered Memory
2022.06生成墙Autoregressive Decode Memory Wall
2023.05KV新墙KV Cache Scaling
2023.09分页KVPagedAttention / Virtualized KV Cache
2024.02带宽墙Asymmetric KV Cache Quantization
2024.04带宽绕道Hierarchical Speculative Decoding
2024.08边界出现Workload-Dependent Bottleneck
2024.10缓存资产化Precomputed KV Reuse
2024.11跨模型共享Cross-Model KV Transfer
2025.07记忆即能力Long-Term Agent Memory
2026.04PIM容量墙Processing-in-Memory Capacity Wall
2026.07可计算的遗忘Rate-Distortion Memory Compaction
未来 6—12 月状态网络Disaggregated Stateful LLM Serving
未来 9—18 月记忆操作系统Agent Memory Operating System
未来 12—24 月内存中心计算Memory-Centric AI Co-Design