ROBCOMM: Robust Compute-in Memory using Memristors

新兴应用（如物联网和大数据分析）对当前的计算机架构和技术构成了严峻的挑战。因此，迫切需要探索替代架构，不仅以更低的成本进一步提高计算效率，而且进一步降低整体能耗。此外，新的器件技术正在出现，与新的架构相结合，可以与不再缩放的CMOS竞争。将计算移动到内存（以数据为中心的计算）是一种新兴的计算范式，在整体计算效率方面具有巨大的潜力。这种计算范例也被称为内存计算（CIM）。将计算转移到内存中（而不是在CPU中进行）将显著减少通信，从而降低功耗并提高性能。CIM架构最有前途的解决方案是基于使用新兴的设备技术，例如电阻或磁阻设备，它们能够充当存储和信息处理单元。这些，更一般地称为忆阻器件，由于器件是非易失性的并且泄漏实际上为零，并且它们也不需要任何刷新，如DRAM等的情况，因此，忆阻器件有利于以较低功耗增加系统复杂性和性能;因此，为科学界提供了新的计算机体系结构创新的机会，能够跟踪今天的限制。然而，实现这样的范例强烈依赖于开发高质量，可靠和节能的电路原语，使存储和计算。本项目旨在设计和展示基于忆阻器件的高质量，可靠，鲁棒和节能的原始电路，以实现存储器中的计算架构。特别是，我们的目标是通过以下方式开发基于忆阻器件的鲁棒CIM电路：i）仔细建模和分析忆阻器件（不）可靠性对CIM操作的影响; ii）开发CIM电路的可靠性感知和容错/容错设计。

Emerging applications (such as Internet-of-Things and Big Data analytics) are posing serious challenges on current computer architectures and technologies. Therefore, there is an urgent need to explore alternative architectures, not only to further increase the computing efficiency at lower cost, but also to further reduce the overall energy. Further, new device technologies are emerging that, in combination with new architectures, may compete with no-longer-scaling CMOS. Moving computation to memory (data-centric computing) is an emerging computing paradigm shown to have a huge potential in terms of overall computing efficiency. This computing paradigm is also referred to as Computation-in-Memory (CIM). Moving the computation to the memory (rather than doing it in the CPU) will significantly reduce the communication and therefore reduce the power consumption and increase the performance.The most promising solutions for CIM architectures are based on the use of emerging device technologies, such as resistive or magneto-resistive devices, that are able to act as both storage and information processing unit. These, more generally called memristive devices, reduce the overall energy consumption as the devices are non-volatile and the leakages is practically zero, and also they do not require any refresh as it is the case for DRAM, etc. Hence, memristive devices favour increasing system complexity and performance at lower power consumption; thus, providing the scientific community with opportunities for new computer architecture innovations being able to track today’s limitation. However, realizing such a paradigm strongly depends on the development of high quality, reliable and power efficient circuit primitives that enable both storage and computing.This project aims at designing and demonstrating high quality, reliable, robust and energy efficient primitive circuits based on memristive devices to enable computation-in-memory architectures. In particular, we aim at the development of robust CIM circuits based on memristive device by i) carefully modeling and analyzing the effects of memristive device (un)reliability on CIM operation and ii) develop reliability-aware and fault/variation tolerant design of CIM circuits.