Tartan: A General-Purpose Spatial Architecture for Exploiting Future Technologies

摘要：Tartan项目的目标是通过理解空间计算中的空间/时间权衡来创建高性能节能计算系统。Tartan消除了ISA的抽象，允许编译器通信管理应用程序执行所需的所有信息。Tartan编译完整的应用程序到应用特定的电路。这些电路可能相当大，并挑战商业CAD工具的可扩展性。在这种情况下，Tartan利用基于全局关键路径的新技术将优化器集中在电路的真正关键元素上。这种新的优化方法适用于整个CAD社区。格子的底层计算基板是一种可重构的织物。这种织物是规则的，因此可以用未来的纳米技术制造。为了提高功率并确保可重构结构能够容忍这些技术中预期的参数变化，结构采用无时钟自定时电路实现。用于在格子中实现这些电路的工具应该具有广泛的适用性，希望能够消除采用自定时电路的一个障碍。当今经济背后的驱动力之一是信息技术的发展，而信息技术本身是由不断增加的可用性和不断降低的计算成本所驱动的。对格子呢的研究将有助于将这些趋势延续到未来。空间计算具有降低设计复杂性、提高性能、提高能源效率和支持未来纳米级技术的潜力。

Abstract:The aim of the Tartan project is to create a high-performance energy-efficient computing system by understanding the space/time trade-off in spatial computing. Tartan eliminates the abstraction of the ISA allowing the compiler to communicate all information necessary to manage the execution of the application. Tartan compiles complete applications into application specific circuit. These circuits can be quite large and challenge the scalability of commercial CAD tools. In this context, Tartan harnesses new techniques based on the global critical path to focus the optimizer on the truly critical elements of the circuit. This new approach to optimization is applicable to the entire CAD community.The underlying computing substrate in Tartan is a reconfigurable fabric. This fabric is regular and thus amenable to manufacturing with future nanoscale technologies. To improve power and to ensure that the reconfigurable fabric is tolerant of the expected parametric variation in such technologies the fabric is implemented using clockless self-timed circuits. The tools used to implement these circuits in Tartan should have broad applicability, hopefully eliminating one the roadblocks towards the adoption of self-timed circuits.One of the driving forces behind today's economy is the growth of information technology, which itself is driven by the constantly increasing availability, and decreasing cost, of computing. Research into Tartan will help to continue these trends well into the future.Spatial computing has the potential to reduce design complexity, improve performance, improve energy efficiency, and support future nanoscale technologies.