CAREER: Architectural Support for CPU / GPU Hybridization

职业：CPU / GPU 混合的架构支持

• 批准号: 1149539
• 负责人: Joseph Zambreno
• 金额: $ 46.63万
• 依托单位: Iowa State University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2012
• 资助国家: 美国
• 起止时间: 2012-04-01 至 2019-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1149539&HistoricalAwards=false
• 关键词: CAREER; Architectural; Support; CPU; GPU

One of the fundamental challenges facing the processor industry today is in continuing to address performance and scalability demands given the constraints of both energy and programmer efficiency. This CAREER project builds upon recently-introduced fused designs, which incorporate conventional microprocessors and graphics processors in the same package, in order to address these challenges through the development of more programmable, flexible, scalable, and secure hybrid architectures. Towards this long-term goal, this project investigates techniques that minimize limitations of the traditional coprocessing model, including performance bottlenecks, vulnerable shared memory spaces, and inflexible resource management. Specifically, this project explores thread divergence caused by complex control-flow and memory access, thread migration between cores, hardware accelerator design for tasks beyond traditional graphics processing, and adding security safeguards to match what is currently available on conventional microprocessors. In addition, this project also examines a variety of instruction set extensions, accelerator architectures, and models of computation that have the potential to lead to significant efficiency and programmer flexibility improvements. The expected impact is that this CAREER project will one day bear fruit in the form of another new segment of the computing industry, one that specializes in hybrid architectures that can scale to thousands of independent threads, while maintaining the capabilities of a traditional multi-user operating system and middleware. Broadly speaking, by advancing the state-of-the-art in computer architecture, this research will lead to improvements in consumer devices in terms of size and form factor, energy efficiency and battery life, and complexity and graphical fidelity of supported applications. Similarly, the usage of these architectures to enable continued performance scaling has the potential to have wide-ranging societal impact, through the advancement of those fields (e.g. bioinformatics, climate modeling, business intelligence) that utilize high-performance computing to enable innovative discovery. This project also includes a focus on expanding an existing course that the PI has developed, that takes a novel approach in introducing graphics processing and architecture from a hardware design perspective.

当今处理器行业面临的基本挑战之一是在能源和程序员效率的限制下继续满足性能和可扩展性需求。该CAREER项目建立在最近推出的融合设计的基础上，该设计将传统微处理器和图形处理器整合在同一封装中，以便通过开发更可编程、灵活、可扩展和安全的混合架构来应对这些挑战。为了实现这一长期目标，该项目研究了最大限度减少传统协处理模型限制的技术，包括性能瓶颈、易受攻击的共享内存空间和不灵活的资源管理。具体来说，该项目探讨了由复杂的控制流和内存访问、内核之间的线程迁移、传统图形处理之外的任务的硬件加速器设计以及添加安全保护措施以匹配传统微处理器当前可用的安全措施所引起的线程分歧。此外，该项目还研究了各种指令集扩展、加速器架构和计算模型，它们有可能显着提高效率和程序员的灵活性。预期的影响是，这个 CAREER 项目有一天将以计算行业的另一个新领域的形式结出硕果，该领域专门研究混合架构，可以扩展到数千个独立线程，同时保持传统多用户的功能操作系统和中间件。从广义上讲，通过推进最先进的计算机体系结构，这项研究将改善消费设备的尺寸和外形、能源效率和电池寿命，以及支持的应用程序的复杂性和图形保真度。同样，通过利用高性能计算实现创新发现的领域（例如生物信息学、气候建模、商业智能）的进步，使用这些架构来实现持续的性能扩展有可能产生广泛的社会影响。该项目还重点扩展 PI 开发的现有课程，该课程采用新颖的方法从硬件设计的角度介绍图形处理和架构。

项目成果

Improving First Level Cache Efficiency for GPUs Using Dynamic Line Protection

使用动态线路保护提高 GPU 的一级缓存效率

• DOI: 10.1145/3225058.3225104
• 发表时间: 2018-08-13
• 期刊: Proceedings of the 47th International Conference on Parallel Processing
• 作者: Xian Zhu;R. Wernsman;Joseph Zambreno
• 通讯作者: Joseph Zambreno

An Adaptive Memory Management Strategy Towards Energy Efficient Machine Inference in Event-Driven Neuromorphic Accelerators

事件驱动神经形态加速器中实现节能机器推理的自适应内存管理策略

• DOI: 10.1109/asap.2019.000-2
• 发表时间: 2019-07-01
• 期刊: 2019 IEEE 30th International Conference on Application-specific Systems, Architectures and Processors (ASAP)
• 作者: Saunak Saha;Henry Duwe;Joseph Zambreno
• 通讯作者: Joseph Zambreno