E2CDA: Type II: Memory, Logic, and Logic in Memory Using Three Terminal Magnetic Tunnel Junctions

现代处理器在阅读内存时消耗的能量至少是执行简单算术时的100倍。这种“内存墙”的起源是处理器和内存速度之间的差距越来越大。系统消耗的功率弥补了支持逻辑（如缓存）的速度差距，并且在等待内存时，未使用的处理器周期会消耗更多的功率。一个潜在的解决方案是使用自旋而不是电荷来存储和处理信息。自旋电子学在逻辑和存储器方面的双重能力将用于解决两项任务中的“存储墙”：首先将存储器写入时间从~ 10 ns减少到1- 10 ps，其次将一些逻辑功能直接纳入存储器阵列。重点将是演示乘法和累加电路，这是所有逻辑系统的基础，但在这里针对卷积神经网络进行了优化。将通过工业研究人员的研究访问以及安排学生实习和最终就业来促进工业联系。教育使命的其他目标是：1）激发K-12学生对科学和工程的职业兴趣; 2）为本科生提供前沿研究机会; 3）通过课程开发和跨学科合作为学生未来的工程职业做好准备。该计划的技术目标将通过使用新技术对合成反铁磁体进行询问和电操作来实现，从而提高磁性隧道结逻辑和存储器的速度。克服集成拓扑绝缘体和磁性绝缘体的材料挑战将减少接近热力学极限的能量耗散。通过拓扑绝缘体中高效的自旋电流产生的低开关能量和合成反铁磁体中的快速开关的组合预计将提供低于1fJ.ps的能量延迟产物。

Modern processors dissipate at least 100× more energy when reading memory than performing simple arithmetic. The origin of this "memory wall" is the increasing gap between processor and memory speeds. Systems dissipate power making up for this speed gap with support logic such as caches, and further power is consumed in unused processor cycles while waiting for memory. One potential solution is to use spin rather than charge to store and process information. The dual capabilities of spintronics in logic and memory will be used to address the "memory wall" in two tasks: first by reducing memory write times from ~10ns to 1-10ps, and second by incorporating some logical functions directly within memory arrays. The focus will be to demonstrate multiply and accumulate circuits the foundation of all logic systems, but optimized here for convolutional neural networks. Industrial connections will be fostered through research visits from industrial researchers and the placement of students for internships and eventual employment. The additional aims of the educational mission are: 1) to inspire career interest of K-12 students in science and engineering; 2) to provide undergraduate students with cutting-edge research opportunities; and 3) to prepare students for future engineering careers through curriculum development and interdisciplinary collaboration. The technical goals of this program will be realized using new techniques for the interrogation and electrical manipulation of synthetic antiferromagnets, thereby enhancing the speed of magnetic tunnel junction logic and memory. Overcoming the material challenges of integrating topological and magnetic insulators will reduce the energy dissipation towards thermodynamic limits. The combination of low switching energy via efficient spin current generation in topological insulators, and fast switching in synthetic antiferromagnets is expected to deliver energy-delay products below 1fJ.ps.