XPS: FULL: Emerging Nonvolatile Memory for Analog-iterative Numerical Methods

一种新型的计算机存储器——交叉点电阻式存储器——已经成为未来计算系统中取代当前存储器技术的一个可能的候选。这种存储器允许潜在的计算机设计具有高存储器容量，并且存储器直接集成到处理单元中。为了开发这些新系统的潜力，需要对计算方法进行新颖的思考。该项目将探索数值优化领域的新基本方法，这些方法适合在包含交叉点电阻性存储器的计算机系统上实现。选择优化领域作为试验台是因为它对广泛的科学学科的重要性。交叉点内存在容量和访问延迟方面具有前所未有的优势。需要大量的创新，以充分利用将内存集成到处理单元的潜在好处；目前的算法是不合适的，因为它们受到冯·诺伊曼瓶颈的约束。pi将设计千兆级模拟迭代网络求解器（gain），这是一种系统架构，可实现高效的原位数据处理。gain改变了应用级、架构级和逻辑/电路级的抽象，使每一层的设计人员和开发人员能够独立工作。(i)它将矩阵提升为一级数据类型。它集成了计算和内存，避免了传统内存层次结构的陷阱。（iii）它利用存储和计算中的多值表示。（iv）用多值模拟电路代替二进制逻辑电路，减少了面积开销和功耗。pi将研究这种变化的范式对数值优化和机器学习算法设计的影响。

A new type of computer memory - crosspoint resistive memory - hasemerged as a likely candidate to replace current memory technology infuture computing systems. This memory allows for potential computerdesigns with high memory capacity and with memory incorporateddirectly into processing units. Novel thinking about computationalmethods is required to exploit the potential of these novelsystems. This project will explore new, fundamental methods in thefield of numerical optimization that are suited to implementation oncomputer systems that incorporate crosspoint resistive memory. Thefield of optimization is chosen as a testbed because of its importanceto a wide range of scientific disciplines.Crosspoint memory has unprecedented advantages in capacity and accesslatency. Substantial innovation is required to fully exploit thepotential benefits of integrating memory into processing units;current algorithms are unsuitable, because they are constrained by thevon Neumann bottleneck. The PIs will design the Gigascale Analog IterativeNetwork Solver (GAINS), a system architecture to enable efficientin-situ data processing. GAINS alters the application-, architecture-,and logic/circuit-level abstractions that enable designers anddevelopers at each layer to work independently. (i) It promotesmatrices to a first-class data type. (ii) It integrates computataionand memory, avoiding pitfalls of conventional memoryhierarchies. (iii) It exploits multi-valued representations in storageand computation. (iv) It replaces binary logic circuits withmulti-valued analog circuits, reducing area overhead and powerconsumption. The PIs will investigate the effects of this changed paradigmon the design of algorithms in numerical optimization and machinelearning.