科研之友
微信
新浪微博
Facebook
分享链接
摘要
The growing number of cores in chip multiprocessors increases the importance of interconnection for overall system performance and energy efficiency. Compared to traditional distributed shared-memory architectures, chip-multiprocessors (CMPs) offer a different set of design constraints and opportunities. As a result, a conventional packet-relay multiprocessor interconnect architecture is a valid, but not necessarily optimal, design point. Worsening wire delays, energy-inefficient routers, and the decreased importance of in-field scalability, make the conventional packet-switched network-on-chip a less attractive option. An alternative solution uses well-engineered transmission lines as communication links. These transmission lines, along with simple, practical circuits using modern complementary metal-oxide-semiconductor technology, can provide low latency, low energy, high throughput channels which can be used as a shared-medium point-to-point link. The design of the transmission lines and transceiver circuits has important architectural impact. This paper includes a first-step design effort for these components, particularly when used for a globally shared-medium bus. For medium-scale CMPs, this interconnect backbone can eliminate the need for packet switching and provide energy, as well as performance benefits when compared to a conventional mesh interconnect. We will provide a design of such a system from the ground up, including design of the transmission lines, transceiver circuits, and a simple, yet effective, architectural design for a shared-medium interconnect, and show that such a design can be a compelling alternative to packet-switched networks for CMPs.
- 出版日期2012-6
- 单位浙江大学
