软件定义专用计算体系结构设计方法研究

The Moore’s Law upholding the evolution of chips generation by generation is facing a grand challenge. By contrast, many diverse new applications featuring super computing intensity quickly spring up, which embarrasses the staggering Moore’s Law. Domain-specific computing architectures are promising to keep delivering the performance demand for specific niche application domains. However, the state-of-the-art design methodology is still not fully explored yet, and a comprehensive systematic co-design paradigm from architecture design to runtime management is still puzzled by many open questions. This project, targeting the domain of computational finance, proposes to use “software-defined” methodology for domain-specific computing architectures. There are two merits of the software-defined methodology: 1) enabling hardware adaptability, which can relax some applicability constraints from specific algorithms, and 2) separating control-flow and data-flow, which can remove the inefficiency of hardware by simplifying the notorious complex control logics. To make the new exploration, we shall delving into three aspects of the project: 1) workload extraction, 2) kernel-based microarchitecture, and 3) programming, complication and runtime system. These research explorations is expected to tackle three fundamental problems dealing with the co-design between throughput and realtime, the co-optimization between control flow and data flow, and the synergy between computation and memory. With the sponsorship, we shall deliver valuable design references for future domain-specific computing architectures. These efforts are expected to spawn a series of original cutting-edge researches in computer architecture and system field.

维持芯片代际更迭的摩尔定律已经面临严峻的挑战。同时，大量以“算力”为基础的应用端的创新层出不穷，这使得芯片发展和计算效能之间的矛盾愈发突出。采用专用计算架构是提升特定应用领域计算效能的有效途径。然而，现有专用计算架构的设计方法还远没有形成统一的定式，也没有形成系统地从结构设计到运行时管理的协同设计方法。本项目面向金融计算这个特定领域，提出通过“软件定义”的方法来设计专用计算架构。我们发现软件定义在专用计算架构上有两个独特的优势：1)赋予硬件一定的“适应性”，有助于解决专用加速器过于“专用”受限于特定算法的不足；2)分离计算中的控制流和数据流，有助于解决复杂控制导致的硬件效率低下的难题。本项目拟从1)金融计算领域负载模式提取，2)专用计算核函数的计算架构设计，和3)编程、编译和运行时系统等三项研究内容展开，拟解决计算通量与实时性的协同、控制流与数据流协同、和计算与存储协同三个关键科学问题。

维持芯片代际更迭的摩尔定律已经面临严峻的挑战。同时，大量以“算力”为基础的应用端的创新层出不穷，这使得芯片发展和计算效能之间的矛盾愈发突出。采用专用计算架构是提升特定应用领域计算效能的有效途径。然而，现有专用计算架构的设计方法还远没有形成统一的定式，也没有形成系统地从结构设计到运行时管理的协同设计方法。本项目面向金融计算这个特定领域，提出通过“软件定义”的方法来设计专用计算架构。软件定义在专用计算架构上有两个独特的优势:1)赋予硬件一定的“适应性”，有助于解决专用加速器过于“专用”受限于特定算法的不足;2)分离计算中的控制流和数据流，有助于解决复杂控制导致的硬件效率低下的难题。.项目全面完成了预定目标，并且在成果转化方面有新的突破。本项目完成金融计算领域的主要计算模式的提炼，并形成一套面向FinTech的基准程序集“FTBench”，并完成在通用CPU、GPU上的性能分析。提出“软件定义加速器体系结构（SDA）”的核心技术方法，并在金融计算领域、数据库计算、深度学习等多领域依据该方法完成专用加速器结构设计的实现，并实现基于FPGA的原型验证系统。这些成果也为后续专用数据处理器DPU的芯片设计做了必要理论准备。