Spin waves to the rescue: Development of a spintronic reservoir computing platform

当我们接近3nm晶体管通道长度的理论极限时，CMOS架构的制造挑战变得更加困难和昂贵。与此同时，计算工作量也发生了翻天覆地的变化，从离线处理转向由物联网（IoT）、机器人和自主代理驱动的实时大数据应用。这些因素的结合导致了对替代计算方法的深入探索，这些方法涵盖了从物理效果到材料、设备、体系结构和数据表示的整个布尔计算堆栈。它还包括新颖的非布尔计算方法，如量子、波和神经形态计算、玻尔兹曼机等。究竟哪一种计算元素的组合将从这众多的选择中进化出来还远不清楚。然而，有可能陈述未来计算平台必须满足的一般需求。首先，任何新的计算方法必须与现有的数万亿英镑的基础设施兼容，这些基础设施与当前基于CMOS的计算相关。其次，它必须通过多代增量硬件和软件改进而可扩展。第三，性能/成本指标必须大大超过布尔CMOS处理器，第四，新技术必须提供比现有技术更节能的替代方案。油藏计算（RC）利用模拟物理系统中的快速非线性动力学，将系统的自发瞬态响应映射到分类任务和信号预测等传统难题的解决方案中。这种技术有效地将记忆和处理任务与材料的固有特性联系起来。然而，只要满足以下关键标准，实现RC的物理系统的具体细节就无关紧要：动态非线性、高相空间维数、唯一可再现的初始状态、易失平衡摄动和动态状态的可读性。主要任务是确定适合任务的系统，该系统不会受到现实世界不兼容需求的困扰。我们提出的解决方案是基于驱动自旋波激励，它既保证了足够复杂的瞬态响应，控制混沌性，又为直接驱动和读取动态磁态提供了一个自然的自旋电子平台。我们提出的工作旨在证明自旋波干扰的多功能性，作为在实际设备中实现RC的关键候选。我们相信磁性纳米结构中的自旋波是开发电路元件和独立器件的嵌入式替代品的理想候选者。这一努力的成功将证明是开发实时模式检测技术的开创性成果，有可能在从医疗监测到气候模拟等领域进行影响深远的部署。在面积为平方微米、功耗为100微瓦、推理时间为10ns的RC硬件上，可以完成复杂的模式识别任务。与目前行业领导者（b谷歌、苹果、Facebook等）为满足全球需求而使用的服务器堆栈相比，这一行动的成功将为大规模、更高效的资源大数据解决方案铺平道路。

As we approach the theoretical limit of 3 nm transistor channel lengths, manufacturing challenges of CMOS architectures become exponentially more difficult and more expensive to overcome. Simultaneously, a seismic shift is occurring in the computational workload, away from offline processing to real-time big-data applications driven by the Internet of Things (IoT), robotics and autonomous agents. This combination of factors has led to an intensified exploration of alternative computing methodologies that span the entire Boolean computational stack from physical effects, to materials, devices, architectures, and data representations. It also includes novel, non-Boolean methods of computing such as quantum, wave and neuromorphic computation, Boltzmann machines and others. Exactly which combination of computational elements will evolve from this plethora of options is far from clear. However, it is possible to state general requirements future computing platforms must meet. First, any new computing methodology must be compatible with the existing multi-trillion-pound infrastructure associated with current CMOS based computing. Second, it must be scalable through multiple generations of incremental hardware and software improvements. Third, the performance/cost metric must greatly exceed that of Boolean CMOS processors, and, fourth, the new technology must provide a much more energy-efficient alternative to existing technology.Reservoir Computing (RC) leverages fast nonlinear dynamics in analogue physical systems to map a system's spontaneous transient response to solutions of traditionally hard problems such as classification tasks and signal prediction. This technique effectively ties memory and processing tasks to the intrinsic materials properties. The specific details of the physical system in which RC is implemented, however, are not relevant so long the following key criteria are met: dynamical non-linearity, high phase space dimensionality, uniquely reproducible initial state, easy out-of-equilibrium perturbation, and readability of dynamical state. The main quest is to identify a system suitable for the task, which is not plagued by real world-incompatible requirements. Our proposed solution is based on driven spin-wave excitations which guarantee both sufficiently complex transient responses, controlled chaoticity, as well as providing a natural spintronic platform for straightforward driving and reading of dynamical magnetic states. Our proposed work aims at demonstrating the versatility of spin-wave interference as the key candidate for the implementation of RC in a real-world device. We believe that spin-waves in magnetic nanostructures are ideal candidates for developing drop-in substitutes for circuit components, as well as stand-alone devices. Success in this endeavour would prove groundbreaking for the development of real-time pattern detection technologies with the potential for high-impact deployment in areas ranging from medical monitoring to climate modelling. Complex pattern recognition tasks could be performed on RC hardware with square-micrometre surface area, 100 micro-W power consumption and 10 ns inference time. Compared to the server stacks currently used by industry leaders (Google, Apple, Facebook, etc.) to satisfy global demand, success in this action will pave the way for massively more resource efficient big-data solutions.