|
马上注册,结交更多好友,享用更多功能,让你轻松玩转社区。
您需要 登录 才可以下载或查看,没有账号?注册
×
Abstract:The Network-on-Chip (NoC) within the Graphics Processing Unit (GPU) is of great significance in GPU design. As GPUs are increasingly being used in diverse fields like artificial intelligence and deep learning, the NoC's performance directly impacts the overall efficiency of these applications. This paper undertakes a comprehensive analysis of the sources and characteristics of the GPU's requirements for the NoC. It delves into the factors such as the high bandwidth demands due to the GPU's multi-threaded, data-intensive concurrent processing characteristics, its unique many-to-few-to-many traffic pattern, and the special cache structure. Subsequently, the paper meticulously summarizes the optimization directions and technical methods at the hardware level. These include various techniques like improving the NoC's topological structure, enhancing the routing algorithms, and optimizing the cache structure in coordination with the NoC. Overall, this paper provides a comprehensive framework for the future research aimed at optimizing the NoC in GPUs, which can lead to more efficient and powerful GPU-based computing systems.
https://ieeexplore.ieee.org/abstract/document/11069468
|
|