This paper presents an in-memory computing (IMC) architecture developed on an 8x8 array of 8T SRAM cells. This architecture enables both multi-bit parallel Multiply-Accumulate (MAC) operations and standard memory processing through charge-sharing on dedicated read bit-lines. By leveraging the maturity of SRAM technology, this work introduces an 8T SRAM-based IMC architecture that decouples read and write paths, thereby overcoming the reliability limitations of prior 6T SRAM designs. A novel analog-to-digital decoding scheme converts the MAC voltage output into digital counts, which are subsequently interpreted to realize fundamental logic functions including AND/NAND, NOR/OR, XOR/XNOR, and 1-bit addition within the same array. Simulated in a 90 nm CMOS process at 1.8 V supply voltage, the proposed design achieves 8-bit MAC and logical operations at a frequency of 142.85 MHz, with a latency of 0.7 ns and energy consumption of 56.56 fJ/bit per MAC operation and throughput of 15.8 M operations/s.
提出了一种紧凑、精确、位宽度可编程的存内计算(IMC)静态随机存取存储器(SRAM)宏(CAP-RAM),用于高效节能的卷积神经网络(CNN)推理。与传统IMC设计相比,它利用了一种新颖的电荷域乘法和累积(MAC)机制和电路,在工艺变化下实现卓越的线性度。采用半并行架构,有效地存储滤波器从多个CNN层共享8个标准的6 T SRAM单元与一个电荷域MAC电路。此外,支持多达六个级别的位宽的权重与两个编码方案和八个级别的输入激活。7位电荷注入SAR(ciSAR)模数转换器(ADC)摆脱了采样保持(S&H)和输入/参考缓冲器,进一步提高了整体能效和吞吐量。一个65纳米的原型验证了良好的线性和计算精度的CAP-RAM。单个512 x128宏存储一个完整的修剪和量化的CNN模型,在MNIST数据集上实现98.8%的推理准确度,在CIFAR-10数据集上实现89.0%的推理准确度,峰值吞吐量为573.4千兆操作每秒(GOPS),能效为49.4兆操作每秒(TOPS)/W。
A 65nm 8b-Activation 8b-Weight SRAM-Based Charge-Domain Computing-in-Memory Macro Using A Fully-Parallel Analog Adder Network and A Single-ADC Interface
Tsinghua University(清华大学);Institute of Microelectronics of the Chinese Academy of Sciences(中国科学院微电子研究所);University of Electronic Science and Technology of China(电子科技大学)
AccelCIM: Systematic Dataflow Exploration for SRAM Compute-in-Memory Accelerator
School of Integrated Circuits, Peking University(北京大学集成电路学院);School of Electronics Engineering and Computer Science, Peking University(北京大学电子工程与计算机科学学院);School of Integrated Circuits, Southeast University(东南大学集成电路学院);School of Integrated Circuits, Anhui Polytechnic University(安徽工业大学集成电路学院)
SMART: Investigating the Impact of Threshold Voltage Suppression in an In-SRAM Multiplication/Accumulation Accelerator for Accuracy Improvement in 65 nm CMOS Technology
CIM-Tuner: Balancing the Compute and Storage Capacity of SRAM-CIM Accelerator via Hardware-mapping Co-exploration
School of Integrated Circuits, Southeast University(东南大学集成电路学院);National Center of Technology Innovation for EDA(EDA国家技术创新中心);Department of Electronic Engineering, Tsinghua University(清华大学电子工程系)
A Digital SRAM-Based Compute-In-Memory Macro for Weight-Stationary Dynamic Matrix Multiplication in Transformer Attention Score Computation
Chongqing University(重庆大学);Institute of Computing Technology, Chinese Academy of Sciences(中国科学院计算技术研究所);Southwest University(西南大学);Institute of Semiconductors, Chinese Academy of Sciences(中国科学院半导体研究所)
DIRC-RAG: Accelerating Edge RAG with Robust High-Density and High-Loading-Bandwidth Digital In-ReRAM Computation
The Hong Kong University of Science and Technology(香港科技大学);University of Southampton(南安普顿大学);South China University of Technology(华南理工大学);Wuhan University(武汉大学)
CAMA: Energy and Memory Efficient Automata Processing in Content-Addressable Memories
加速有限自动机处理对于推进模式匹配、数据采掘、生物信息学、入侵检测和机器学习中的实时分析至关重要。最近利用SRAM和DRAM的内存中自动机加速器已经显示出比传统数字设计令人兴奋的改进。然而,所有SOTA设计所使用的状态转换的位向量表示仅在处理最坏情况的完全随机模式时是最佳的,而在运行大多数真实世界的基准测试时会浪费大量的内存和能量。我们提出了CAMA,内容可寻址存储器(CAM)启用自动机加速器处理同质非确定性有限自动机(NFA)。一个完全不同的状态表示方案,以及共同设计的新电路和数据编码方案,大大降低了能源,内存和芯片面积最现实的NFA。本文对CAMA进行了全面优化,主要贡献如下:(1)采用16 × 256的8管(8 T)CAM阵列进行状态匹配,取代SOTA设计中的256 × 256的6 T SRAM阵列或两个16 × 256的6 T SRAM组;(2)提出了一种新的编码方案,能够在8 T SRAM中进行内容搜索,并适应不同的应用;(3)可重新配置和可扩展的架构,在所有测试基准上提高效率,而不会失去对与SOTA设计兼容的任何NFA的支持;(4)优化框架,自动选择编码方案并将给定的NFA映射到所提出的硬件。两个版本的CAMA,一个优化的能量(CAMA-E)和其他的吞吐量(CAMA-T),在28纳米CMOS工艺,并在21个真实世界和合成基准进行了全面评估。CAMA-E的能量比CA、2-…