Analog content-addressable memory from complementary FeFETs
Analog content-addressable memory from complementary FeFETs
复制标题
来自互补 FeFET 的模拟内容寻址存储器
DOI：
10.1016/j.device.2023.100218
复制
发表时间：
2024
期刊：
影响因子：
--
通讯作者：
Jariwala, Deep
中科院分区：
文献类型：
--
作者：
Liu, Xiwen;Katti, Keshava;He, Yunfei;Jacob, Paul;Richter, Claudia;Schroeder, Uwe;Kurinec, Santosh;Chaudhari, Pratik;Jariwala, Deep
尽管最近在用于矩阵乘法的非易失性存储器（NVM）方面取得了进展，但其他关键的数据密集型操作（如并行搜索）仍然在很大程度上被忽视。当前的并行搜索架构，即内容可寻址存储器（CAM），通常使用二进制，这限制了密度和功能。我们提出了一个模拟CAM（ACAM）单元，建立在两个互补的铁电场效应晶体管（FeFET），在模拟域中进行并行搜索超过40个不同的匹配窗口。ACAM不仅提供了比三元CAM（TCAM）更密集3倍的内存架构，而且对于使用Omniglot数据集模拟的少次学习，在相似性搜索方面的推理准确性提高了5%，与基于硅基互补金属氧化物半导体（CMOS）节点。此外，我们在ACAM中演示了对内核回归模型的一步推理，模拟结果表明推理速度比CPU和GPU快1，000倍。
Despite recent advancements in non-volatile memory (NVM) for matrix multiplication, other critical data-intensive operations like parallel search remain largely overlooked. Current parallel search architectures, namely content-addressable memory (CAM), often use binary, which restricts density and functionality. We present an analog CAM (ACAM) cell, built on two complementary ferroelectric field-effect transistors (FeFETs), that performs parallel search in the analog domain with over 40 distinct match windows. ACAM not only offers a projected 3× denser memory architecture than ternary CAM (TCAM) but also yields a 5% increase in inference accuracy on similarity search for few-shot learning simulated with the Omniglot dataset, with an estimated speedup per similarity search of more than 100× when compared to a central processing unit (CPU) and graphics processing unit (GPU) on scaled silicon-based complementary metal-oxide semiconductor (CMOS) nodes. Furthermore, we demonstrate one-step inference on a kernel regression model in ACAM, with simulation results indicating 1,000× faster inference than a CPU and GPU.