SHF: Small: Acceleration Using Smart Memory-on-Chip

片上存储器结构几乎用于所有现代计算机处理芯片，包括通用芯片和加速器芯片。它们占据了芯片面积的很大一部分，例如，在英特尔的服务器级至强处理器中，仅用于其最后一级缓存的空间就有35 MB，而谷歌的Tensor处理单元则有24 MB用于片上存储。此外，与实际计算相比，处理器花费不成比例的大部分时间和能量在其存储器层次结构上移动数据以及在指令处理中。为了解决这些效率低下的问题，该项目提出了一个新的想法：重新利用内存结构中的元素，并将它们转换为大型数据并行计算单元。这项研究的结果有可能加速涉及大量使用数据集的计算操作。存储在片上存储器阵列中的数据共享导线（位线）和信号感测装置（senseamps）。该研究是在obeservation的基础上，算术运算可以计算在这些共享的结构，通过增加一些门。这种SRAM内计算技术被称为位线计算。初步工作表明了智能存储器芯片的潜力和可行性。该项目将探索新颖的垂直集成解决方案，探索智能芯片存储器的广泛使用。该研究将开发智能存储器的新操作原语、编程框架和编译器，设计神经计算架构，研究智能存储器在专用集成电路（ASIC）和可重构现场可编程门阵列（FPGA）中的应用，该奖项反映了NSF的法定使命，并通过使用基金会的学术价值和更广泛的影响审查标准。

Memory-on-Chip structures are used in almost all modern computer processing chips, both general-purpose as well as accelerators chips. They occupy a large fraction of the die area, for example, in Intel's server class Xeon processor devotes 35 MB just for its last-level cache, while Google's Tensor Processing Unit dedicates 24 MB for on-chip storage. Furthermore, a processor spends disproportionately large fraction of time and energy in moving data over its memory hierarchy, and in instruction processing, as compared to actual computation. To tackle these inefficiencies, this project proposes a novel idea: re-purpose the elements in memory structures and transform them into large data-parallel compute units. Outcomes of this research have the potential to accelerate computing operations involving heavy use of data sets.Data stored in on-chip memory arrays share wires (bit-lines) and signal sensing apparatus (senseamps). The research is grounded in the obeservation that arithmetic operations can be computed over these shared structures by augmenting a few gates to them. This in-SRAM computing technique is referred to as bit line computing. The preliminary work has demonstrated the potential and feasibility of smart memory-on-chip. This project will explore novel vertically integrated solutions that explore broad use of smart memory-on-chip. The research will develop new operation primitives, programming framework and compiler for smart memories, design neural computing architectures, investigate utilizing smart memories in Application Specific Integrated Circuits (ASICs) and reconfigurable Field Programmable Gate Arrays (FPGAs), and explore the applicability to emerging embedded nonvolatile technologies.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.