Resonant Compute-In-Memory (rCIM) 10T SRAM Macro for Boolean Logic
Resonant Compute-In-Memory (rCIM) 10T SRAM Macro for Boolean Logic
复制标题
DOI：
10.1109/iccd58817.2023.00026
复制
发表时间：
2023-11
期刊：
影响因子：
--
通讯作者：
Dhandeep Challagundla;Ignatius-In Bezzam;Biprangshu Saha;Riadul Islam
中科院分区：
文献类型：
--
作者：
Dhandeep Challagundla;Ignatius-In Bezzam;Biprangshu Saha;Riadul Islam
传统的最先进的计算平台依赖于硅基静态随机存取存储器(SRAM)和数字布尔逻辑来进行密集的计算。尽管金属氧化物半导体晶体管已经得到了积极的扩展，但基本的冯-诺伊曼计算体系结构仍然没有改变。内存计算(CIM)的出现为克服传统冯-诺依曼体系结构中的存储墙瓶颈提供了一种很有前途的解决方案，它允许在SRAM存储元件中处理和存储信息。本文介绍了一种能量回收共振式10T-SRAM体系结构，该体系结构促进了内存中的计算，以最大限度地减少处理核心和内存之间的数据移动需求。利用串联谐振写驱动器在写操作期间有效地回收放电能量，以降低SRAM体系结构的整体能耗。通过在使用TSMC28nmPDK的8KB存储阵列上实现，验证了所提出的rCIM的可行性。此外，还进行了全面的蒙特卡罗变化分析，以确保方案在工艺变化下的稳健性和可靠性。为了验证该体系结构的有效性，我们使用EPFL组合基准测试套件对其性能进行了评估。与标准冯-诺伊曼结构相比，该结构的功耗降低了55.42%，吞吐量达到88.2GOPS/S。
Traditional State-of-the-Art computing platforms have relied on silicon-based static random access memories (SRAM) and digital Boolean logic for intensive computations. Although the metal-oxide-semiconductor transistors have been aggressively scaled, the fundamental von-Neumann computing architecture has remained unaltered. The emerging paradigm of Compute-in-Memory (CIM) offers a promising solution to overcome the memory wall bottleneck in traditional von-Neumann architectures by enabling the processing and storing of information within SRAM memory elements. This article introduces an energy-recycling resonant 10T-SRAM architecture that facilitates in-memory computations to minimize the need for data movement between the processing core and memory. Series resonant write driver is utilized to efficiently recycle the discharged energy during a writing operation to reduce the overall energy consumption of the SRAM architecture. The feasibility of the proposed rCIM has been demonstrated by implementing it on an 8KB memory array using TSMC 28nm PDK. Additionally, a comprehensive Monte Carlo variation analysis was conducted to ensure the robustness and reliability of the scheme under process variations. To demonstrate the effectiveness of the proposed architecture, we evaluate its performance using the EPFL combinational benchmark suite. The proposed Resonant Compute-In-Memory (rCIM) consumes 55.42% lower energy than standard von-Neumann architecture and achieves a throughput of 88.2-106.6 GOPS/s.