{
 "sources": [
  "https://arxiv.org/abs/2012.03112",
  "https://semiengineering.com/is-in-memory-compute-still-alive/",
  "https://doi.org/10.1109/MSSC.2019.2922889",
  "https://ieeexplore.ieee.org/abstract/document/9512855",
  "https://doi.org/10.1038/s41565-020-0655-z"
 ],
 "about": "骨架：一级按「算在哪」（存内 / 近存，Mutlu《A Modern Primer on Processing in Memory》、Semiconductor Engineering《Is In-Memory Compute Still Alive?》2024-12），二级按模拟 / 数字（Verma 等 IEEE SSC Magazine 2019），存储介质作字段（Shimeng Yu 等 IEEE CAS Magazine 2021；Sebastian 等 Nature Nanotechnology 2020）。outside_main_axis = 存算主轴之外只列代表的框（研究线、相邻路线、上游）。",
 "layers": [
  {
   "code": "a1",
   "name": "RRAM / 忆阻器",
   "desc": "权重写成阻变器件的电导，交叉阵列里一步算完矩阵乘",
   "box": "模拟存内计算",
   "box_desc": "权重存在存储单元里，用电流 / 电荷直接做乘加",
   "outside_main_axis": false
  },
  {
   "code": "a2",
   "name": "闪存 / NOR",
   "desc": "用成熟的 Flash 单元存权重，做模拟乘加",
   "box": "模拟存内计算",
   "box_desc": "权重存在存储单元里，用电流 / 电荷直接做乘加",
   "outside_main_axis": false
  },
  {
   "code": "a3",
   "name": "电荷域 / 电容及其他模拟",
   "desc": "用电容、忆容器或全模拟电路算",
   "box": "模拟存内计算",
   "box_desc": "权重存在存储单元里，用电流 / 电荷直接做乘加",
   "outside_main_axis": false
  },
  {
   "code": "d1",
   "name": "端侧 / 边缘",
   "desc": "多为 SRAM 存算，做摄像头、汽车、PC 上的推理",
   "box": "数字存内计算",
   "box_desc": "存储阵列里嵌数字乘加，精度和良率跟普通数字芯片一样",
   "outside_main_axis": false
  },
  {
   "code": "d2",
   "name": "数据中心推理",
   "desc": "存内计算 + chiplet / 3D 堆叠，专做大模型推理",
   "box": "数字存内计算",
   "box_desc": "存储阵列里嵌数字乘加，精度和良率跟普通数字芯片一样",
   "outside_main_axis": false
  },
  {
   "code": "n1",
   "name": "DRAM / HBM 近存（PIM）",
   "desc": "内存条、HBM 或 3D 堆叠 DRAM 里放计算单元",
   "box": "近存计算",
   "box_desc": "计算放在存储旁边或存储芯片里，减少搬数据",
   "outside_main_axis": false
  },
  {
   "code": "n2",
   "name": "端侧近存（at-memory）",
   "desc": "计算单元紧贴片上存储的低功耗端侧芯片",
   "box": "近存计算",
   "box_desc": "计算放在存储旁边或存储芯片里，减少搬数据",
   "outside_main_axis": false
  },
  {
   "code": "n3",
   "name": "SRAM 近存大芯片",
   "desc": "相邻代表：靠大容量片上 SRAM 换带宽，不算真存内",
   "box": "近存计算",
   "box_desc": "计算放在存储旁边或存储芯片里，减少搬数据",
   "outside_main_axis": false
  },
  {
   "code": "r1",
   "name": "研究线",
   "desc": "IBM、TSMC、Samsung、imec、高校的存算研究芯片",
   "box": "大厂与学界研究线",
   "box_desc": "还没变成商品的存算芯片：论文和研究原型",
   "outside_main_axis": true
  },
  {
   "code": "x1",
   "name": "光子计算",
   "desc": "用光做矩阵乘或芯片间互连",
   "box": "相邻路线",
   "box_desc": "也想绕开「存储墙」的另两条路，只列代表",
   "outside_main_axis": true
  },
  {
   "code": "x2",
   "name": "类脑 / 神经形态",
   "desc": "脉冲、事件驱动，存和算在同一个「神经元」里",
   "box": "相邻路线",
   "box_desc": "也想绕开「存储墙」的另两条路，只列代表",
   "outside_main_axis": true
  },
  {
   "code": "u1",
   "name": "新型存储 IP 与器件",
   "desc": "RRAM / MRAM / FeFET 等存储 IP 和器件",
   "box": "上游",
   "box_desc": "存算芯片要用到的新型存储 IP、器件、设备和工具",
   "outside_main_axis": true
  },
  {
   "code": "u2",
   "name": "设备、材料与 EDA",
   "desc": "新型存储的沉积设备、材料和设计工具",
   "box": "上游",
   "box_desc": "存算芯片要用到的新型存储 IP、器件、设备和工具",
   "outside_main_axis": true
  }
 ]
}